If you were wondering the same thing I am - it's not about skills loss, quality, and less about money spent. It's more about frontier AI shops dogfooding their own models.
If it’s anything like AWS there’s hundreds of people making bespoke software factory setups, enhanced interfaces for ai tools, spinning up 10 parallel review agents with the best model available, etc because the budget is basically unlimited.
The other part that is kind of comical is that the software factories keep growing quality gates and automated tasks that need to run for every commit, PR, deploy, etc. Then someone goes "oh, but now it's slowing us down", so a new thing gets added which decides when it's appropriate to run the gate, and on and on endlessly until it's a gigantic soup of actions that are running which provide negligible, and more frequently negative value over the software lifecycle. People are making big expensive messes of agents and acting like it's galaxy brain stuff.
I find it a bit comical. AI tools are so easy to pick up, and it's unlikely you've pushed some highly critical features which couldn't have waited for a few months.
You make a good point. That’s why there is AI fatigue. People are moving at speeds that aren’t normal and it’s mentally taxing.
If it’s not already a thing, it will be soon, the mental health aspects of all professions moving at unsustainable speeds and what that will do to people.
Budget unlimited + upper management claiming that engineers should be spending at least their salary equivalent on token costs or they're being ineffective. So many possible misaligned incentives...
Microsoft is not a frontier AI shop at all. They just resell others' models. They don't have anything of their own worth mentioning. They missed the boat and tried to acquire OpenAI and failed, and now they pivoted their strategy to be a model-agnostic middleman.
And meta is only arguably so. Not really in the same league as OpenAI and Anthropic. More second tier like Google and xAi. (and of those Google is pretty close to the top two at times)
Talking to my friends working there, they said they were surprised when that happened and they didn't find Claude quality to be that much better than Gemini when using agy. So maybe after all the issue is not mainly the model, but harness and custom tooling.
I dunno, it could be about money, too. $100,000 budget per employee, per month. Why were they willing to spend that much on AI usage? I hope said employees also make that much in salary.
I think if you talk to LLMs and give feedback or openly say what works and what doesn't, you are essentially solving a captcha and produce accurate training data, while you pay for the token spend. I'd be a bit nervous with this lol.
Just one unsanitized input and you leak info. Or one hidden character and code may or may not belong to you anymore. Its very odd on many levels
These are companies famous for following the rules when it comes to handling other people’s data and IP after all, totally a non-issue and they would never violate contract law
Phi was more of a research project. Not trained on real world data but synthetic input. That's really good to control if you go for strict limits but it also loses a lot of originality.
Only if all you care about is developing models. I assume the rest of the business would rather just use whatever's best in class regardless of who made it, so I'm sure it's not that straightforward of a decision.
I can't be sure but I vaguely recall it being associated with the big Microsoft antitrust court case? Along with the delightful phrase "knifing the baby"
> Within Microsoft’s cloud and AI sector, monthly AI spending limits have reportedly been slashed from $100,000 per employee to approximately $10,000 in the majority of instances.
They were allowing $100,000 per month per employee?!
That's incredible. Just a few years ago that would have been zero.
I'm curious if they have any data showing whether it benefits productivity. I get by on the free tier of ChatGPT, and do 100% of the coding myself. It has never been a bottleneck, and I like to keep my skills sharp. AI helps in analyzing problems, getting quick answers to technical queries, reviewing code, etc. But running my own business I'm incredibly careful about expenses.
My company took away my Claude because it’s too expensive. I feel like there is a reckoning coming. The accountants are finally realising the cost of token maxing.
That's pretty stupid. Most people who are incurring significant costs are just tokenmaxxing rather than being efficient with usage. You can get 99% of jobs and work done with Haiku/Luna in a collaberating working enviroment.
I feel like people who are later to the AI game just like to "oneshot" and sink a bunch of usage into generating garbage
There really is a skill to using it effectively. I've tried coaching some of the devs on my team. Some get it, some don't.
Our company has been tracking token usage and models used vs output (tickets, story points, PRs, deploys, etc...). A dev got chewed out, even after I warned him, because he spent over $2k in a single month almost exclusively on Opus while his actual productivity in terms of what he delivered was abysmal.
I get it but it goes against the grain for me. Isn't it ironic that we have to waste our precious and expensive human brain cycles to think about how to use AI cheaply so that it is not more expensive than us?
In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
> In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
It sounds like the parent is less talking about this, and more people burning tokens while not getting useful work done.
Claude, vibe code me an entire startup, the actual product doesn't matter, but it should all be based on the incredible pun "turn 'sorry' points into story points."
/goal get accepted into Y Combinator, you have an unlimited token budget, be bold.
EDIT: no, do not just make a product that gives away your unlimited token budget to users for free!
Attempting a serious but not-a-certified-whatever answer: "Points" do have meaning when properly used as a kind of moving-average tool for forecasting within a particular context.
Problems arise when people try to perma-peg them to particular tasks, or (worse) man-hours or (much worse) man-hours across teams. Even just encouraging the humans to answer in terms of hours/days taints the accuracy of the forecast by introducing a kind of bias.
So, what you do is you recognize every ticket has a somewhat variable “actual effort”; and, if you’ve been honest in approximate effort pointing, you’ll know your team (or your own) velocity.
From there you can run Monte Carlo simulations - say a few hundred thousand, and get a pretty good estimate of actual time spent.
Assume that for various practical reasons, we've decided it's one of the best functions out there. How do we use it effectively, especially when it has noise, and drifts over time with unseen changes to the human_estimator and the hideously complex world_state?
A popular option is to run it multiple times with different person/task combinations, putting a projected number on to each task. Afterwards, the tasks finished in sampling period ("sprint") become a quantifiable total for that period ("velocity").
Do the same process again with the next set of tasks, and you can figure out which ones are likely to fit if the velocity doesn't change much. If you know the velocity will change due to losing staff or vacation days... well, we apply a multiplier and hope for the best.
Trying to "fix" the meaning of points is maladaptive, because they reflect many changing things which are outside our control and can't be independently measured.
My wife, despite loving all things French, just doesn't "do metric". She wants my height in feet and inches, my weight in pounds, boom done. So while I know my mass in kilograms, she needs the conversion done before she can even begin to have a reference point.
Upper management is the same way. They have forecasts that they need to make, deadlines and budget goals that they need to hit. They only deal in the units of hours and dollars (or local currency). Every software engineer is accountable for their work in those units only. The conversion needs to be done before the management chain has a reference point.
One easy way to do this is to have each engineer estimate the time it takes to fulfill a story after it's been pointed; then, upon completion, record their actual hours spent. Their estimated vs. actuals tend to stabilize over time, so even if they misestimate a task, you can arrive at a good guess at the time it will actually take.
> > Problems arise when people try to perma-peg [points to man-hours]
> Upper management [...] only deal in the units of hours
I feel you're mixing up different operations here. You can always express unfinished work as likely to require a certain number of team-sprints, which are convertible to theoretic man-hours. The key is that the conversation rate is only valid for a moment, and technically that moment was the prior sprint.
That's very different from management thinking (or worse, declaring) that points have a permanently fixed proportion to man-hours.
> [...] and dollars
If your management deals in international currency, then perhaps that would be a useful analogy to them: Points and Man-Hours are different sides of FOREX, and they fluctuate based on different conditions.
When Engineering predicts a group of tasks is 54 points, that's like a foreign company signing a long-term contract in €100 EUR instead of USD. You can estimate that you'll receive ~$112 USD in a year, but the actual dollars will likely be different because the exchange-rate will continue changing before that happens.
It's funny, the thing that makes effective prompt also makes effective documentation/communication.
It's bizzare to see people that made clown issues (not enough detail etc.) suddenly start writing detailed prompts just because it is AI that will do the task and not the human on the other side.
The guys that made clown issues are still using AI to make clown issues, they just look like plausible specs now that claude wrote it more thoroughly. At least pre claude I could tell that they had missed something earlier on, ask a question to clarify, and get a real answer they had to type themselves. Now most of my stories are rehashed after I raise PRs or even after these same guys approve the PR and then realize they forgot a requirement.
You mean that now rather than taking 30 seconds to write a ticket, because we all know what we're talking about, we must take 10 minutes to give all the background information to the AI?
there's a manifold to what "effective" means. The problem is once you get into the vibe flow, it's really difficult to eject yourself into the other realms of vscode or IDE or whatever it is you normal do because the vibing provides no anchor to what you're doing.
Even if these models are smart enough to reorient themselves, they get entirely stuck in a desert and now you're asking someone to just pull up stakes and digg them out even thought they only watched them get there and the UI provides so much speed that no human can comprehend how they got there in the first place.
It's like asking a pilot to take over in an emergency situation when they're not tasked with any of the every day requirements of the job. The orgs are relying on borrowed time of experienced professionals, and that's going to erode away and what replaces it is mostly people who understand how to navigate context but not use any of the _classic_ tools.
It's a real conundrum and won't be easily surfaced but for a decade.
I’m trying really hard to keep my skills up but it doesn’t feel productive when I’m using it to write code. It feels like I’m slowing down the AI to the point that it’s not as effective as just letting it go. But I don’t get all the learning that comes from that time along the way.
Have you found ways to stay sharp while using it? Or are you relying on other projects outside of work to keep your skills fresh?
I'm primarily using local models. I turn on thinking visible and try to review at speed what it's doing.
Other than that, no. I drag myself out to poke around occasionally because at times the local models get to far into context and refuse to do simple tasks.
I'm not exactly one-shotting software, but because of the pace I'm expected to keep, I place a lot more trust in the bot that I'm actually comfortable with. I need the job for the time being, so I just keep hitting the button and letting it do its thing until the tests are green.
I'm just working in DevOps though, so it's writing IaC, not application code (save the odd Lambda function or python script). Still, even when I spend an entire day conversing with Claude and watching "bot go brrrrr", I'm one of the lowest users in our company. I have no idea what the devs who regularly hit their limits are doing.
I feel like it's only within the past few months that opus got to the point where guiding the model is faster than doing things myself. I tried out sonnet recently and it was not a net positive to my work. I feel like anything that I'd trust haiku to handle isn't worth doing in the first place.
For context, I'm doing a range of tasks, everything from one-shotting adhoc scripts to having 4 hour 10M+ token conversations debugging things.
I’m a different person and definitely not a genius but my experience today goes even beyond theirs.
I had Opus trying to simplify a query for me which was slow - it ran for maybe 30 minutes, including writing and running tests, and came up with a refactor across 9 files with a couple hundred lines changed. I was looking through the output before moving onto the next step, and noticed something a little fishy- I said “why does it do x, isn’t that a more complex y?”
Opus thought for another 20-30 seconds then output “Actually that would make the majority of the diff irrelevant, if we do that change it is just these 4 lines in this single file instead.
So then I had it do that. 5-10 minutes of writing and testing and that was done.
So my company spent $25 in tokens and I spent probably an hour in total for a 4 line change that, in the days before Claude, I probably could have found the correct file and thought through the problem, understood the solution, and written the 4 lines of code myself. Probably in the same amount of time.
So basically there was no benefit at all for my time, an extra cost to the company of $25, and now I understand our codebase a little bit less instead of more if I had done all the work.
As good as Claude is at building greenfield projects it still struggles a lot at complex ones
This is a recurring issue for us. Very small team. We move fast and pretty loose.
Dev + AI spend 3-4 hours on a project plan, there's a "wait a minute" moment, and finally they spend another hour dialing it back to a solution that could have been built, tested, and deployed in 2 hours.
Example: Someone was setting up a dev environment with multiple DB migrations from different branches - AI planned this wild 8 phase solution with a pretty fancy cutover event.
In review I essentially said... "Wait, isn't this a dev environment? It doesn't need 0 downtime, why not just destroy and recreate the DB" and it turned into a <1000LOC script.
Technically the original plan would have worked, it would have been more robust, but it would have taken a good deal more time to implement.
Some of this falls on the devs to know what fits our team well, what's realistic, what's obviously overengineered, etc... But some of it feels like AI just defaults to the most complex version of a thing. I catch it SUPER frequently. (And unfortunately some devs think that more complexity means it's a better solution)
It's not so much that I can code fast, it's that it takes a significant amount of time to tell the model what exactly I want the script to accomplish, and at that point I might as well write the script myself. And often, figuring out what I want the script to do is that hardest part, so it doesn't really matter whether writing the script takes 10% or 20% of the total time.
in my company there are a few who keep sharing screenshots of reaching limits on 3 separate subscriptions, 2 of them their personal on top of the company subscription
Wonder when subscription-hopping attacks will become more often (jumping from a personal model to injecting instructions into the business account and exfiling data)
I think they mean the new Haiku, which is mildly above Luna now . If you have a plan written by a smarter model (so the hard parts are solved) they can be great at implementation.
It's impressive for its total size if you need local inference.
Though it's significantly slower in Token/s and also thinks a lot more without matching the same intelligence (xhigh qwen27b scores lower than haiku's medium setting, and haiku-med is $0.05 per task compared to Qwen27B's $1.01 on AA's comparison)
It seems like more a backup if you need to work offline, imo, unless time doesn't matter and/or your electricity is free. Or you want independence from the labs (fair enough).
Or does a "collaborating work environment" mean that everything is basically spoonfed to them? Or do you only ever use ghost suggestions?
I genuinely cannot even fathom. Just how do you even get into a state where tasks are so clear and cookie cutter? These things are abhorrent. Not only are they not useful, it's an outright form of psychological torture to try and use them. They almost fight you.
Luna doesn't even respond to steers properly! You try steering it and it immediately gets distracted and then just stops.
I can imagine coercing Sonnet into doing some of my tasks okay, but Haiku? Especially 4.5? Really?
I think you might be overestimating the sort of projects most of us have worked on throughout our careers -- we haven't been doing much groundbreaking work. LLMs can easily and successfully write most code.
It's possible it's my role distorting my perception, cause technically I don't write software, I work an SRE role. None of my items come pre-chewed or paced, it's all good luck and god bless.
I'm desperately trying to classify and standardize my work items and delegate them to less capable models, because my usage is clearly unsustainable and this same sentiment as above keeps being pushed on me too. But all my tasks are genuinely fairly arbitrary, so there's no real way around the agent actually being able to reason about business and technical context proper. It's not even that they're hard, it's just that they're dynamic.
I can get Luna to do things like walk our observability stack and perform a healthcheck, then defer to a stronger model if anything looks super off, but if I'm being entirely honest, this could basically be just a script. Which Opus 5.5 will immediately write for itself if it doesn't yet exist, run that, and then off it goes depending. But Luna will never actually do an investigation proper. Heck, it can't even read our dashboards most of the time, tripping up on Grafana minutia.
It feels like that surgeon vs surgeon comparison, where you're made to decide based on their surgery success rate, and the better succeeding surgeon simply reward hacks the number by only operating on less dicey cases. Except there's no objective way to make this classification here, so jackasses like the above get to play with my insecurities with full obnoxious confidence, while I'm left desperately trying to slim my usage and failing to do so between two moments of crippling self doubt and blockers.
For the last few weeks I've been using Luna with Medium Reasoning for routine debugging, installing/updating dependencies, test creation/fixes, CLI/config miscellaneous problem solving, etc and it's been solid. I can let it churn away for a half hour and it barely moves the needle on remaining usage on my Plus plan. Previously I had been using Sol Low/Medium and would frequently hit the 5 hour limit doing those kind of tedious automation and routine problem solving tasks.
I wish we had such a platform (and was properly adopted). Maybe then the necessary context would be properly organized, and these lesser models could be effective here as well, especially if combined with harnessing integrations too. I still have a hard time accepting that Haiku/Luna tier models can be effective there even then, but I'll just have to take your word for it I suppose.
That's funny. In a meeting with my manager recently, they specifically called me out for not spending enough money on tokens. It's not like I didn't use it, just apparently not enough, and apparently on too low a setting.
Since then I've had fable cranked up to 11 for even the most trivial of tasks.
I know plenty of people working at very well known and very large companies who's CEOs were boasting about "not hiring anyone anymore", "all the code will be generated in 3 months", etc. they all went from "unlimited budget per dev" + public dashboard with ranking to flex how much credits everyone was burning to hard caps at $500-$1000/month/employee real fast. Some are even not allowing their devs to use the more expensive models
Do you know the details of the Claude Code plan you and your company are using (if not part of some enterprise deal)? Does your individual capacity out run something like Claude Max 20x ($200/mo)?
I've got a 100$ openAI sub through work and I never even get close to reaching the limit. I really wonder what the hell is everyone else doing that they could even get close to spending 4 figures a month in tokens.
As a point of reference, I tried an experiment with Claude Code and the latest Opus the other day. It was work for my company, so this was using API tokens and not the individual user plans that have non-commercial terms.
A simple task, migrating a typical password reset flow as part of updating a long-lived web application from legacy libraries and software architecture to modern equivalents, apparently cost roughly the same as 2 months of Pro subscription, over the equivalent of about half a working day in wall time.
It produced code of decent quality at a small scale, but it wasn’t always on point architecturally. It also had a tendency to drift off topic and try to tangle up other changes it decided should be made with the main change we were supposed to be working towards. So even for a routine task, based on a plan developed using the harness first and with the agents working under close supervision, a near-SOTA model is still producing quality on par with a decent mid-level developer but substandard for anyone senior+ in this case.
Moreover, based on a direct comparison with other migration tasks of similar complexity that I’d already done by hand, it was actually a bit slower overall to work this way. I had to babysit Claude throughout and review everything it proposed carefully, both to avoid subtle errors (it would have made several) and to prevent drifting off track. I also had to spend a significant amount of time cleaning up its final output to an acceptable standard after the session. Those two overheads more than cancelled out the much faster code generation an LLM offers under favourable conditions.
So for now, I remain sceptical about these high multiples of improved productivity that I keep seeing claimed online from people who are apparently writing almost everything using AIs now. I could certainly have achieved a multiple of my normal productivity by YOLOing everything without reviewing it in detail and then accepting the output code without tidying anything up. However, I doubt this codebase would still have been good enough for normal human developers to work on it reasonably after even 10 or 20 AI-led sessions like that. The architecture would have degraded significantly and the test suite would have been large and largely pointless. And again, this wasn’t rocket science in this experiment, it was completely unremarkable maintenance of a relatively small and simple web application.
I use Opus 5.5 heavily but only spend around $800/week at API rates. I mean, I say "only"... That's a lot in absolute terms, but trivial compared to my salary and EASY worth it.
And someday I hope to understand why they do that. CEO: "Let's see, I can pay $200/month for Bob's tokens, or I can pay $2000/month or so, and then hope he doesn't screw up and rack up a seven-figure bill. The service is the same either way. Hmm."
Is there something in the ToS that says you can't give each of your employees a personal $200/mo subscription... and if they happen to use it at work, well, that's their business?
Even if so, it might be worth spinning up a separate LLC for each division, given what they are charging for the API.
Does all this velocity translate to increased income to the business though. At some point if we are releasing 20x more features, more games, more music, who is actually buying it all?
At this point we're solidifying a lot of stuff for the product: disaster recovery, getting workflows to scale so we can actually sell it to more people, paying off the Fort Knox of tech debt, so, at least in my case, yeah. YMMV.
I'm in a similar boat. When I joined my current company the tech was this 700KLOC legacy behemoth stuck in 2005. It was pretty much all tech debt.
With agents I've been able to make a decent sized dent in it. Lots of this gain is due to AI - but the fact of the matter is we still have 50+ bigger projects we could work on, and non-developers building with AI has increased that number.
Is that not a widely-known word in project management circles for scrum and other methodologies? I've heard it for years. Basically how much work you get done through the lens of how long it took to get done via a pointing system.
I didn't know people were using these non ironically.
Might as well go back to counting the number of lines or number of commits. It doesn't sound as good as "granularity" and "velocity" though. The good thing with velocity is that it doesn't care which way you're going as long as you're going there fast, so you can never be wrong
The reckoning started years ago when we did the equivalent to token maxxing hiring coders for everything to crank LOC
Software is inherently a physics problem not all the job titles and specializations made up the last 20 years as dev job salaries kept attracting people
That was all illusory social construct to prop up jobs
Still a whole lot of that in tech but it's all at the top of the org now. Leadership sensory experience and thus innate habit to forecast future been programmed by years of yes men they refuse to accept the jig is up for them too
Sensory memory of being a useless figurehead fosters a lot of existential dread in priests, politicians, and the like. Completely aware their day to day effort is insufficient to sustain them they know how co-dependent they are. They'll dig in harder.
See Chris Matthews flame out shrieking about socialist execution squads. Dude seriously thought everyone wants to hang him from a lamp post. The reality is people just want a sense of control back and not have their perception dragged along by Chris Matthews.
which is quite sad because opus 5.5 is really good. i say this as an anthropic hater. i wish I could move away to other models like 6.1 sol or deepseek or whatever, but they just all lack something. i _trust_ opus 5.5
i hope other labs catch up, especially chinese labs.
I mean, we’re not far from a situation where instead of how many story points you completed per sprint the metric to optimize is going to be what was your efficiency? How many story points did you complete while minimizing your token usage. In fact, that’s a pretty good idea. I’m going try and implement it at work with some sort of complexity normalization function
If true, it is a huge blow to Anthropic’s revenue stream. IIRC it was reported that the quarter of their revenue comes from just two clients and as the ex-Meta guy who left this July, I am convinced that Meta must be one of the two.
> The company's financial trajectory already shows how quickly that equation is changing. Anthropic's revenue run rate was about $9 billion at the end of 2025, according to the company, before rising to more than $47 billion by May. Anthropic has projected revenue of at least $10.9 billion for the second quarter of 2026, more than double the previous quarter, on track for its first quarterly operating profit of $559 million.
> The company has told a small group of shareholders that its adjusted operating income will be positive for the second consecutive quarter, according to multiple people with knowledge of the matter.
These are leaked and self-reported numbers, but no matter how much skepticism you pile on them it still looks likely that Anthropic in 2026 have had some of the fastest revenue growth of any company in history.
They may be self-reported numbers, but they're consistent across multiple reporting sources. If Anthropic are lying to their investors about these numbers they will be in very real trouble with the SEC come IPO time.
And even if they're using funky accounting tricks to exaggerate their profits and downplay some of their losses, I expect the numbers for 2026 will still be really impressive.
I'll check w/ archive.ph in a bit, thanks for the tip.
Is the hope for them that they'll outpace their run with a massive revenue growth? I believe that they have that, but numbers like a net loss of $46b in 2025[0] seem hard to overcome.
Re: self-reporting, I'd be inclined to think they're using non-GAAP practices to make it sound more favorable, but I am by no means an expert, so willing to believe I am wrong.
I'll read the article you sent; this is just my off-the-cuff thinking.
They claim to have been profitable in both Q2 and Q3 (under whatever their definition of profitability is).
I really don't think the 2025 figures are interesting at this point. Everything changed for them in 2026 - nobody was spending $1000/month/employee in 2025, there wasn't enough interesting token-heavy stuff to do with the models.
I think Anthropic might actually make it to profitability - they're earning more revenue than OpenAI and they've spent significantly less, too.
Doesnt matter if Microsoft and meta are pushing employees towards their own models today. What matters is that it's the economically sensible direction, and so it will happen if it's not today.
There is a conspiracy theory that the tokenmaxxing period was anthropic/openai ipo play. They do the tokenmaxxing to revenuemaxxing first, then time the ipo window to show they had the the huge growth to justify their price tag.
However Elon was able to ipo before them which took a lot liquidity of the market. Their financials are exposed. The market condition and sentiment now is in the gutter. It would be very interesting to see how these would pan out
That's my theory, which I've seen being echoed here a few times. Encourage increased usage, suddenly raise prices, capture the photo-finish in the short lapse between prices going up and usages adjusting, and present that goldilock point-in-time to the investors as the new normal.
Considering that’s a healthy portion of a salary for an additional employee per person, the fact they’re slashing spending sure makes it look like AI wasn’t even a 2x multiplier at minimum.
You could only conclude that if they would have dropped LLMs completely. It just means they see the benefit/cost optimum at a lower point than 100k/programmer.
I also doubt a longterm 2x multiplier for most developers.
I don't think you need the LLMs to be completely dropped to indicate this. Even entry-level faang programmer salaries are in the 300k range, so the fact that LLMs aren't worth 100k/programmer does tell us something about the marginal usefulness. If the drop was motivated purely by price-per-output, it would indicate the LLMs are less useful than adding another 1/3 engineer. However, as discussed in other comments, there are other motivations as well, notably dogfooding homegrown tooling.
Even if it was a net negative, you still might want to titrate the exit a bit just to prevent the total chaos of broken workflows. Can't really draw the conclusions you want to draw.
We're sitting on a year of more or less capable coding models and I have not exactly seen a revolution in new software being released. AI is likely a force multiplier for specific subsets of individuals and workflows. Coding is not a bottleneck for a lot of systems.
I suspect that SaaS is silently being eaten alive as an industry.
I think SaaS will be fine. How the field works might change, but in the end marginal cost of software is still zero and when you've got a quality, well-tested and well-designed product it's going to be obviously better than things vibed together.
You often pay for regulatory compliance and liability with SaaS. Would you risk running your own vibe-coded payroll or accounting system for your company?
Can you link to some of the huge leaps in emulators and binary decompilation that have caught your attention? (I hope that doesn't sound like a challenge. I'm just interested.)
It’s mostly from people using their personal accounts to run LLM services that serve a larger team or organization. At least at Microsoft, it’s still impressively hard to get access to an LLM for service usage with high enough rate limits to be useful, making running services on dev boxes much more appealing (despite the countless drawbacks that few people seem to care about around security, compliance, reliability, etc).
Right! it seems obvious why: Both these companies want to dogfood their own coding models and stop paying competition.
You can also read this as diminishing returns / AI isn't good enough, etc, but the simplest explanation is that they don't want to send money to Anthropic.
This, and in addition to dogfooding, incentivizing employees to be more effective with the cheaper models. A lot of problems don't need anything fancy, but it takes more brain power and engineering effort to make that work. By default humans will take the path of least resistance if it's available.
But Microsoft and Meta are not blocking competitor tools for internal use, they're merely trying to reduce costs and divert a fraction of use to their own technologies. Microsoft and Meta are both still spending nine figures a year on Claude, and the article does not state or imply they're even considering a complete halt.
I feel like "good enough" was reached around Opus 4.6 - 4.8. All I wanted after that is improved speed, continued tweaks to the tooling to get the most out of it and quality of life features added.
The first things I install on Windows 11 are OpenShell and ExplorerPatcher to get rid of their crap taskbar and start menu and bring back the Windows 10 taskbar and start menu. What they did with Windows 11 is an abomination.
What does it even mean to use "Copilot" these days?
I use the Github Copilot app at work but it's really just a tool that gives models access to our Github repos and my local clones of them, and it's been super helpful to just tell it to reference other repos for patterns or integrations or whatever, get agents and an orchestrator to look at multiple repos to make a change across several projects and then implement them, etc.
And yeah you can use a selection of whatever model, I was using Opus 4.6 for a while, now I'm on GPT-6 Sol. It's been a very good experience.
Just to sprinkle more details here about MSFT. The usage was trimmed down even more for some orgs and some specific accounts (I think those were vendors and the like). But if you are an FTE and somewhere in CoreAndAI then your limit is like 1M GitHub credits ($10k as article mentions) that could be spent on any model you have access to, e.g. Opus 5x or GPT-x. It is hard to pin down how many people use Claude vs how many just use internal harness over the Opus though. Importantly, Anthropic models are more expensive (not even talking about Fable here as it is a special case) as seen in usage metrics. Peeps can check their usage which has a good breakdown that shows how many tokens were burned on which model and how much that "costs" (the prices as you can imagine are not consumer ones), so IMO this change will drive more people to optimize based on their usage patterns if they are heavy hitters. The easiest path is to just switch to something like Sol/Astra and have similar perf for lower price.
> Peeps can check their usage which has a good breakdown
Must be nice. At Sony we use Copilot (as an MS enterprise customer) and restrictions to token usage circa June have resulted in a hunger games of who-can-use-them-all-first, with the pool usually drying up by the 10th day of each month.
As for usage breakdown: we got a manually-published table most days just showing per-user token usage, until eventually we were told even that was somehow too difficult to provide.
Between that and learning how much the whole of Sony Corporate Group spends on Copilot (spoiler: it's hilariously little), I've sort of run out of disbelief.
I'm so glad I work at a company that isn't looking at our token usage. I mean, we're encouraged to use it but hearing about leaderboards and things is wild - but we also aren't big enough to need API costs, so power users get a flat rate 200/Month subscription or whatever.
I could see finance looking at it from a cost perspective if it was API costs, but sharing it with the rest of the users feels like a different measurement...
> Importantly, Anthropic models are more expensive (not even talking about Fable here as it is a special case) as seen in usage metrics.
Not sure where you see that. The API prices of Opus are cheaper than Astra and there was a study suggesting that the Anthropic abos are offering more bang for the buck:
But what exactly does Microsoft have to show for this spend? Engineers are allowed to spend up to 100k a month in AI credits. And what comes of that?
Is most of it spent on internal administrative tools?
Is Microsoft cutting contracts with other SaaS vendors because their employees use AI to build their own alternatives?
Is it just everyone querying the same questions to cut through the corporate drudgery and documentation hell?
As an ignorant outsider, Windows is still terrible. Github is very mid. VS Code is... there. SQL Server/Windows server is ... still there
What about Meta? Instagram is the same as its been for years. WhatsApp the same. Meta RayBan app the same. Facebook.com the same. I don't use Snap, so is Anthropic's 2nd largest customer improving Snapchat? I doubt it.
AFAIK there is no such explicit, company-wide effort at meta. The article seems to try to slip Meta in there with whatever reporting they are doing on Microsoft, despite there being no such evidence for meta.
To also add an important context to the drop in Claude code users reported for meta - this elides that we are absolutely still using the Anthropic models full steam ahead, but are moving towards internal interfaces and harnesses.
So I would question if that 50% drop in CC users is more of an interface change than anything else.
I for one have stopped explicitly using codex or CC entirely. But the interfaces I am using still use those harnesses under the hood. I wonder how that is counted.
From what I could see these numbers would usually be hit by people doing vast amounts of analysis over files, they would usually rock up tokens very fast. Then you combine this with the selection of Opus models which are much more expensive internally than GPT ones and the limit gets reached quite quickly.
I can't speak to MSFT, but this is sort of a non-story for Meta. Of course they want employees using their own model! They haven't actually taken any steps to limit Claude access.
I assume that at shops that both employ engineers and are developing an AI product, internal usage is not about improving productivity, it is about improving the offering. Of course they want employees to use internal tools.
Maybe I am slow here and everyone is using Claude with credits at max use. But isn't Claude Teams like $25/month per developer for ordinary use? What the heck of these guys doing that makes it get that phenomenally expensive for their use cases? These are presumably well capable engineers who started to use this as an aid right not just throw Fable at everything and loop to the max?
There are people out there building AI building orchestrators for orchestrators for orchestrators for agents. The author of that blog post later claimed to be spending the equivalent of $122k/month on tokens (by rotating their usage between 21 accounts).
As far as I can tell, the only thing that this level of spend has produced so far is an indie 2D RPG video game.
> Gas Town was intended to be reusable, but I only ever wound up using it to build itself. Gas Town fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly. With 4.7 we saw the introduction of the "just two more things" tic, which prevented Opus from ever converging on being ready to do real work—it always wanted to fiddle with Gas Town itself. The Opus tic never went away, so Gas Town effectively burned down. It had other problems, too, but 4.7 was the final straw
I hadn't seen the $122K figure mentioned previously. $87K for API-style pricing was mentioned in the above post, and ~$2,800/mo for multiple Claude Max accounts:
> My solution has been to create a token tap on $200 Max accounts, which for me work out to ~30x the list-price equivalent. So in reality I'm only spending about $2800/month out of pocket for my $87k "worth" of tokens. Though that number keeps growing alarmingly.
(He never seemed to provide numbers for Gas Town initially, whether what he paid, or what the API-style pricing would have charged, so it was interesting to get some actual figures. Sounds like a lot to me, but if he's genuinely getting multiple people's-worth of work out of it, then...)
(Also in the article: a little morsel of Emacs content, which was nice to see.)
Oh, you can see his outputs. GasTown and Beads (and maybe other AI-related projects) are all open source on GitHub, and you should go checkout the activity on those repos. At their peak he was churning out absolutely insane amounts of code. I'm talking 100s of commits and literally 100s of 1000s of lines of code changed per week.
Whether that produced anything worthwhile can be controversial, depending on the very subjective ways people quantify these things. All I'll say is that they have 18K and 27.7K stars on GitHub and apparently a comparable number of users.
He's always come across to me as at least a little bit like this! His AI stuff doesn't feel enormously out of character to me. (If you hadn't encountered him before Gas Town - he's been posting online for like 20+ years, got a Wikipedia page and everything, so you can find his back catalogue and judge for yourself.)
The cybersecuritynews.com news one simply republishes details of a story published by The Information. At least they have the decency to LINK to that Information story:
> In the landmark United States v. Microsoft Corp. antitrust case, evidence and allegations revealed that Microsoft maintained an illegal operating system monopoly by withholding undocumented or "secret" Windows application programming interfaces (APIs) from third-party competitors while letting its own internal product teams use them for an unfair advantage.
I wonder if these still exist, and if so, is MSFT worried about leaking them if they let their devs use other vendor's models.
I should probably know more about Claude's TOS, but it is probably a mistake for these companies not to leverage the plausible deniability of their usage and turn it into a massive distillation resource for their own models.
At Microsoft, the recommendation is to use cheaper models for tasks that don't need frontier models. I personally use Lua and it's more than capable for complex problems.
It doesn't say they're cutting back on OpenAI usage.
To me this seems mostly related to the way Anthropic showed the level at which they monitor sessions, plus wanting to limit how much training data they're directly feeding into a company that competes with their own products/investments.
CEO's nephew showed him how good the Chinese models are?
I am only half joking, I heard something like "my son or nephew did this cool thing with $X so we'll take $this_radical_step because of it" enough times over my career.
There is a perceived opportunity cost from someone using a lower-tier model on their task. What if the better model did a "better" job? what if my trials and tribulations are due to model quality?
If you are used to talking to opus5.5 medium, going to GPT6.1 luna low will feel like a step down. Why would any employee take the (personal) risk?
It may bei cost efficient, but is it wise? We use not only the big US models, but also Chinese ones. This way we can compare who makes the difference. Simplified: Knowledge comes before economic aspects.
You're opening yourself up to data right and privacy risks with that. My company demands that I use their enterprise account because they can claim full ownership of all produced output and have full logs of every interaction.
I think that gets legally murky, if the employee is the one who pays for the tool.
Wait, okay maybe I'm in a bubble, but every org I've worked at understood that giving us whatever hardware we want more than pays for itself, and very quickly.
I can't fathom the logic of paying 10k a month for a developer and giving them an 800 device.
Insane. Your leadership is severely broken. If you can, work somewhere that values you.
If you were wondering the same thing I am - it's not about skills loss, quality, and less about money spent. It's more about frontier AI shops dogfooding their own models.
If it’s anything like AWS there’s hundreds of people making bespoke software factory setups, enhanced interfaces for ai tools, spinning up 10 parallel review agents with the best model available, etc because the budget is basically unlimited.
The other part that is kind of comical is that the software factories keep growing quality gates and automated tasks that need to run for every commit, PR, deploy, etc. Then someone goes "oh, but now it's slowing us down", so a new thing gets added which decides when it's appropriate to run the gate, and on and on endlessly until it's a gigantic soup of actions that are running which provide negligible, and more frequently negative value over the software lifecycle. People are making big expensive messes of agents and acting like it's galaxy brain stuff.
Ye’ Olde Pendulum be swingin’ back again
If the money is free, wht not.
I find it a bit comical. AI tools are so easy to pick up, and it's unlikely you've pushed some highly critical features which couldn't have waited for a few months.
You make a good point. That’s why there is AI fatigue. People are moving at speeds that aren’t normal and it’s mentally taxing.
If it’s not already a thing, it will be soon, the mental health aspects of all professions moving at unsustainable speeds and what that will do to people.
Budget unlimited + upper management claiming that engineers should be spending at least their salary equivalent on token costs or they're being ineffective. So many possible misaligned incentives...
Token maxing really is a thing if your token budget is unconstrained.
No, because not all work is productive
Tokenmaxxing doesn't say anything about productivity. It's only about using as many tokens as possible; there's nothing about using them usefully.
Tokenmaxxing is just the new pr-maxxing. Proxies truly are the root of all evil.
That sounds both very fun and very stressful.
Microsoft is not a frontier AI shop at all. They just resell others' models. They don't have anything of their own worth mentioning. They missed the boat and tried to acquire OpenAI and failed, and now they pivoted their strategy to be a model-agnostic middleman.
And meta is only arguably so. Not really in the same league as OpenAI and Anthropic. More second tier like Google and xAi. (and of those Google is pretty close to the top two at times)
Meanwhile https://www.reddit.com/r/GeminiAI/comments/1wh0qxq/google_fi...
This made me laugh https://www.reddit.com/r/GeminiAI/comments/1wh0qxq/comment/p...
Talking to my friends working there, they said they were surprised when that happened and they didn't find Claude quality to be that much better than Gemini when using agy. So maybe after all the issue is not mainly the model, but harness and custom tooling.
They give claude usage on antigravity and aren't them a investor on antrhopic?
What in the world is happening?
Ask Claude
We broke up, ask it yourself.
I wonder if there's a couple (2 humans) out there where his name is Claude and her name is Alexa... Or Siri.
Life uhhh... finds a way.
I dunno, it could be about money, too. $100,000 budget per employee, per month. Why were they willing to spend that much on AI usage? I hope said employees also make that much in salary.
it's circular revenue. they can claim tokenmaxing on one side, and huge revenues on the other.
I think if you talk to LLMs and give feedback or openly say what works and what doesn't, you are essentially solving a captcha and produce accurate training data, while you pay for the token spend. I'd be a bit nervous with this lol.
Just one unsanitized input and you leak info. Or one hidden character and code may or may not belong to you anymore. Its very odd on many levels
Accurate training data?
At best you produce some noisy signals that are going to have a tiny impact if even that.
And that's on a personal plan where you didn't opt out of sharing usage data.
Business plans offer zero data retention. This is a non-issue.
These are companies famous for following the rules when it comes to handling other people’s data and IP after all, totally a non-issue and they would never violate contract law
It would be a low quality training set that you theorize is a goldmine and worth illegally stealing from your customers.
I think this data is likely worthless compared to curated RL tasks.
Microsoft?
Copilot
That's not a model.
True, but its not pure OpenAI GPT. If the point is dogfooding, then they'd use Copilot instead of using OpenAI's models directly
They do.
They had phi for a while. Interesting that they haven't really continued with anything like that.
Phi was more of a research project. Not trained on real world data but synthetic input. That's really good to control if you go for strict limits but it also loses a lot of originality.
But it was never a SOTA model.
If you ask Microsoft, it is a lifestyle.
Dogfooding their own model ls and not letting their competitors use their data to train their.
I wonder why they allow it at all.
Like its a no brainer to force your employees to use your own models, then RL train them to be better.
Only if all you care about is developing models. I assume the rest of the business would rather just use whatever's best in class regardless of who made it, so I'm sure it's not that straightforward of a decision.
Or let them use Claude, track everything, and use that data to train your own models.
If you assume that noisy general usage data enables good RL, particularly compared to curated RL training sets.
I am not convinced that's the case.
all of the above. plus less likely to be ripped off by competitor. (yes I know they have all been ripping off the general public)
if AI never happened there's like zero chance I would've ever used or noticed the usage of the word "dogfooding" lmao i hate this timeline
That's odd. I've been aware of that word for decades.
Ditto, it's been around for a decade or three, especially if we include longer phrase "eating your own dogfood" and not just the verbification.
I can't be sure but I vaguely recall it being associated with the big Microsoft antitrust court case? Along with the delightful phrase "knifing the baby"
Same as "load-bearing", but I guess that depends on everyone's non/pre-swe background.
Or house flipping TV shows
Wikipedia has the introduction of the term "dogfood" to the corporate world as a Microsoft internal memo....
... in 1988.
https://en.wikipedia.org/wiki/Eating_your_own_dog_food
MS was famous for that. Seems since around .NET they have forgotten about it. Delivering and promoting toolkits to devs they don't use themselves.
anyone who works in software knows this term
> Within Microsoft’s cloud and AI sector, monthly AI spending limits have reportedly been slashed from $100,000 per employee to approximately $10,000 in the majority of instances.
They were allowing $100,000 per month per employee?!
That's incredible. Just a few years ago that would have been zero.
I'm curious if they have any data showing whether it benefits productivity. I get by on the free tier of ChatGPT, and do 100% of the coding myself. It has never been a bottleneck, and I like to keep my skills sharp. AI helps in analyzing problems, getting quick answers to technical queries, reviewing code, etc. But running my own business I'm incredibly careful about expenses.
This is so ridiculous lol. not paying 20 bucks for a sub has to be a handicap
Just a cool million+ annual expense per headcount.
Notably, the limit per employee != actual average usage per employee!
Microsoft bloat really knows no bounds.
My company took away my Claude because it’s too expensive. I feel like there is a reckoning coming. The accountants are finally realising the cost of token maxing.
That's pretty stupid. Most people who are incurring significant costs are just tokenmaxxing rather than being efficient with usage. You can get 99% of jobs and work done with Haiku/Luna in a collaberating working enviroment.
I feel like people who are later to the AI game just like to "oneshot" and sink a bunch of usage into generating garbage
There really is a skill to using it effectively. I've tried coaching some of the devs on my team. Some get it, some don't.
Our company has been tracking token usage and models used vs output (tickets, story points, PRs, deploys, etc...). A dev got chewed out, even after I warned him, because he spent over $2k in a single month almost exclusively on Opus while his actual productivity in terms of what he delivered was abysmal.
I get it but it goes against the grain for me. Isn't it ironic that we have to waste our precious and expensive human brain cycles to think about how to use AI cheaply so that it is not more expensive than us?
In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
> In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
It sounds like the parent is less talking about this, and more people burning tokens while not getting useful work done.
I dunno, using tools and resources effectively is arguably the essence of good engineering.
This is much like how devs got grilled for creating expensive test VMs on AWS. Someone has to pay for all this at the end of the day.
What do sorry points mean anymore.
I love the typo.
Claude, vibe code me an entire startup, the actual product doesn't matter, but it should all be based on the incredible pun "turn 'sorry' points into story points."
/goal get accepted into Y Combinator, you have an unlimited token budget, be bold.
EDIT: no, do not just make a product that gives away your unlimited token budget to users for free!
Ah not to worry, you'll make it up in volume!
Did they ever have meaning? It's always been a nebulous feels term
Attempting a serious but not-a-certified-whatever answer: "Points" do have meaning when properly used as a kind of moving-average tool for forecasting within a particular context.
Problems arise when people try to perma-peg them to particular tasks, or (worse) man-hours or (much worse) man-hours across teams. Even just encouraging the humans to answer in terms of hours/days taints the accuracy of the forecast by introducing a kind of bias.
For forecasting what, if not man-hours?
Effort. Which is a very nebulous term, I agree.
So, what you do is you recognize every ticket has a somewhat variable “actual effort”; and, if you’ve been honest in approximate effort pointing, you’ll know your team (or your own) velocity.
From there you can run Monte Carlo simulations - say a few hundred thousand, and get a pretty good estimate of actual time spent.
I’ve seen it work before with shocking accuracy.
To play with the math analogies, imagine a black-box function:
estimate(human_estimator, task_description, world_state) -> numeric_effort
Assume that for various practical reasons, we've decided it's one of the best functions out there. How do we use it effectively, especially when it has noise, and drifts over time with unseen changes to the human_estimator and the hideously complex world_state?
A popular option is to run it multiple times with different person/task combinations, putting a projected number on to each task. Afterwards, the tasks finished in sampling period ("sprint") become a quantifiable total for that period ("velocity").
Do the same process again with the next set of tasks, and you can figure out which ones are likely to fit if the velocity doesn't change much. If you know the velocity will change due to losing staff or vacation days... well, we apply a multiplier and hope for the best.
Trying to "fix" the meaning of points is maladaptive, because they reflect many changing things which are outside our control and can't be independently measured.
Sorry.
My wife, despite loving all things French, just doesn't "do metric". She wants my height in feet and inches, my weight in pounds, boom done. So while I know my mass in kilograms, she needs the conversion done before she can even begin to have a reference point.
Upper management is the same way. They have forecasts that they need to make, deadlines and budget goals that they need to hit. They only deal in the units of hours and dollars (or local currency). Every software engineer is accountable for their work in those units only. The conversion needs to be done before the management chain has a reference point.
One easy way to do this is to have each engineer estimate the time it takes to fulfill a story after it's been pointed; then, upon completion, record their actual hours spent. Their estimated vs. actuals tend to stabilize over time, so even if they misestimate a task, you can arrive at a good guess at the time it will actually take.
> > Problems arise when people try to perma-peg [points to man-hours]
> Upper management [...] only deal in the units of hours
I feel you're mixing up different operations here. You can always express unfinished work as likely to require a certain number of team-sprints, which are convertible to theoretic man-hours. The key is that the conversation rate is only valid for a moment, and technically that moment was the prior sprint.
That's very different from management thinking (or worse, declaring) that points have a permanently fixed proportion to man-hours.
> [...] and dollars
If your management deals in international currency, then perhaps that would be a useful analogy to them: Points and Man-Hours are different sides of FOREX, and they fluctuate based on different conditions.
When Engineering predicts a group of tasks is 54 points, that's like a foreign company signing a long-term contract in €100 EUR instead of USD. You can estimate that you'll receive ~$112 USD in a year, but the actual dollars will likely be different because the exchange-rate will continue changing before that happens.
rookie numbers. in one of the top companies, i know someone who tokenmaxed so hard they ended up spending $50000
So you are going to blame this one that dev?
Define productivity, and while at it, quality, maintainability , modularity and so forth.
What are the biggest pitfalls devs on your team fall into?
It's funny, the thing that makes effective prompt also makes effective documentation/communication.
It's bizzare to see people that made clown issues (not enough detail etc.) suddenly start writing detailed prompts just because it is AI that will do the task and not the human on the other side.
The guys that made clown issues are still using AI to make clown issues, they just look like plausible specs now that claude wrote it more thoroughly. At least pre claude I could tell that they had missed something earlier on, ask a question to clarify, and get a real answer they had to type themselves. Now most of my stories are rehashed after I raise PRs or even after these same guys approve the PR and then realize they forgot a requirement.
You mean that now rather than taking 30 seconds to write a ticket, because we all know what we're talking about, we must take 10 minutes to give all the background information to the AI?
there's a manifold to what "effective" means. The problem is once you get into the vibe flow, it's really difficult to eject yourself into the other realms of vscode or IDE or whatever it is you normal do because the vibing provides no anchor to what you're doing.
Even if these models are smart enough to reorient themselves, they get entirely stuck in a desert and now you're asking someone to just pull up stakes and digg them out even thought they only watched them get there and the UI provides so much speed that no human can comprehend how they got there in the first place.
It's like asking a pilot to take over in an emergency situation when they're not tasked with any of the every day requirements of the job. The orgs are relying on borrowed time of experienced professionals, and that's going to erode away and what replaces it is mostly people who understand how to navigate context but not use any of the _classic_ tools.
It's a real conundrum and won't be easily surfaced but for a decade.
I’m trying really hard to keep my skills up but it doesn’t feel productive when I’m using it to write code. It feels like I’m slowing down the AI to the point that it’s not as effective as just letting it go. But I don’t get all the learning that comes from that time along the way.
Have you found ways to stay sharp while using it? Or are you relying on other projects outside of work to keep your skills fresh?
I'm primarily using local models. I turn on thinking visible and try to review at speed what it's doing.
Other than that, no. I drag myself out to poke around occasionally because at times the local models get to far into context and refuse to do simple tasks.
I'm not exactly one-shotting software, but because of the pace I'm expected to keep, I place a lot more trust in the bot that I'm actually comfortable with. I need the job for the time being, so I just keep hitting the button and letting it do its thing until the tests are green.
I'm just working in DevOps though, so it's writing IaC, not application code (save the odd Lambda function or python script). Still, even when I spend an entire day conversing with Claude and watching "bot go brrrrr", I'm one of the lowest users in our company. I have no idea what the devs who regularly hit their limits are doing.
That’s just not how any of this works.
You’re conceptualizing. In reality, AI is expensive, power consuming, planet destroying, and overall productivity killing.
That's what happens when token usage becomes a performance metric. As has been done at my company.
I feel like it's only within the past few months that opus got to the point where guiding the model is faster than doing things myself. I tried out sonnet recently and it was not a net positive to my work. I feel like anything that I'd trust haiku to handle isn't worth doing in the first place.
For context, I'm doing a range of tasks, everything from one-shotting adhoc scripts to having 4 hour 10M+ token conversations debugging things.
That is always amazing to me.
There is no way I can beat even local models at generating complex Python scripts fast.
hn is filled with uber geniuses.
I’m a different person and definitely not a genius but my experience today goes even beyond theirs.
I had Opus trying to simplify a query for me which was slow - it ran for maybe 30 minutes, including writing and running tests, and came up with a refactor across 9 files with a couple hundred lines changed. I was looking through the output before moving onto the next step, and noticed something a little fishy- I said “why does it do x, isn’t that a more complex y?”
Opus thought for another 20-30 seconds then output “Actually that would make the majority of the diff irrelevant, if we do that change it is just these 4 lines in this single file instead.
So then I had it do that. 5-10 minutes of writing and testing and that was done.
So my company spent $25 in tokens and I spent probably an hour in total for a 4 line change that, in the days before Claude, I probably could have found the correct file and thought through the problem, understood the solution, and written the 4 lines of code myself. Probably in the same amount of time.
So basically there was no benefit at all for my time, an extra cost to the company of $25, and now I understand our codebase a little bit less instead of more if I had done all the work.
As good as Claude is at building greenfield projects it still struggles a lot at complex ones
This is a recurring issue for us. Very small team. We move fast and pretty loose.
Dev + AI spend 3-4 hours on a project plan, there's a "wait a minute" moment, and finally they spend another hour dialing it back to a solution that could have been built, tested, and deployed in 2 hours.
Example: Someone was setting up a dev environment with multiple DB migrations from different branches - AI planned this wild 8 phase solution with a pretty fancy cutover event.
In review I essentially said... "Wait, isn't this a dev environment? It doesn't need 0 downtime, why not just destroy and recreate the DB" and it turned into a <1000LOC script.
Technically the original plan would have worked, it would have been more robust, but it would have taken a good deal more time to implement.
Some of this falls on the devs to know what fits our team well, what's realistic, what's obviously overengineered, etc... But some of it feels like AI just defaults to the most complex version of a thing. I catch it SUPER frequently. (And unfortunately some devs think that more complexity means it's a better solution)
Matches my experience to a tee.
And don’t forget the company also spent a bunch of money in tokens for the initial author to implement the thing poorly.
It's not so much that I can code fast, it's that it takes a significant amount of time to tell the model what exactly I want the script to accomplish, and at that point I might as well write the script myself. And often, figuring out what I want the script to do is that hardest part, so it doesn't really matter whether writing the script takes 10% or 20% of the total time.
in my company there are a few who keep sharing screenshots of reaching limits on 3 separate subscriptions, 2 of them their personal on top of the company subscription
Wonder when subscription-hopping attacks will become more often (jumping from a personal model to injecting instructions into the business account and exfiling data)
> You can get 99% of jobs and work done with Haiku/Luna in a collaberating working enviroment.
You can get the same jobs and work done with Kimi and GLM (ZDR on OpenRouter) for a fraction of the price too.
Sorry, but that is nonsense. Compared to opus haiku doesn't cut it most of the time.
I think they mean the new Haiku, which is mildly above Luna now . If you have a plan written by a smarter model (so the hard parts are solved) they can be great at implementation.
I use Qwen3.8-27B as a daily driver, and for things I know will be quite hard I tend to get ChatGPT to do the planning. Works very well.
It's impressive for its total size if you need local inference.
Though it's significantly slower in Token/s and also thinks a lot more without matching the same intelligence (xhigh qwen27b scores lower than haiku's medium setting, and haiku-med is $0.05 per task compared to Qwen27B's $1.01 on AA's comparison)
It seems like more a backup if you need to work offline, imo, unless time doesn't matter and/or your electricity is free. Or you want independence from the labs (fair enough).
What on earth do you even do with these models?
Or does a "collaborating work environment" mean that everything is basically spoonfed to them? Or do you only ever use ghost suggestions?
I genuinely cannot even fathom. Just how do you even get into a state where tasks are so clear and cookie cutter? These things are abhorrent. Not only are they not useful, it's an outright form of psychological torture to try and use them. They almost fight you.
Luna doesn't even respond to steers properly! You try steering it and it immediately gets distracted and then just stops.
I can imagine coercing Sonnet into doing some of my tasks okay, but Haiku? Especially 4.5? Really?
I think you might be overestimating the sort of projects most of us have worked on throughout our careers -- we haven't been doing much groundbreaking work. LLMs can easily and successfully write most code.
I equate most LLM work to squeezing a glue bottle
It's just glue code
It's not complicated. Someone just has to be there to squeeze the bottle
It's possible it's my role distorting my perception, cause technically I don't write software, I work an SRE role. None of my items come pre-chewed or paced, it's all good luck and god bless.
I'm desperately trying to classify and standardize my work items and delegate them to less capable models, because my usage is clearly unsustainable and this same sentiment as above keeps being pushed on me too. But all my tasks are genuinely fairly arbitrary, so there's no real way around the agent actually being able to reason about business and technical context proper. It's not even that they're hard, it's just that they're dynamic.
I can get Luna to do things like walk our observability stack and perform a healthcheck, then defer to a stronger model if anything looks super off, but if I'm being entirely honest, this could basically be just a script. Which Opus 5.5 will immediately write for itself if it doesn't yet exist, run that, and then off it goes depending. But Luna will never actually do an investigation proper. Heck, it can't even read our dashboards most of the time, tripping up on Grafana minutia.
It feels like that surgeon vs surgeon comparison, where you're made to decide based on their surgery success rate, and the better succeeding surgeon simply reward hacks the number by only operating on less dicey cases. Except there's no objective way to make this classification here, so jackasses like the above get to play with my insecurities with full obnoxious confidence, while I'm left desperately trying to slim my usage and failing to do so between two moments of crippling self doubt and blockers.
For the last few weeks I've been using Luna with Medium Reasoning for routine debugging, installing/updating dependencies, test creation/fixes, CLI/config miscellaneous problem solving, etc and it's been solid. I can let it churn away for a half hour and it barely moves the needle on remaining usage on my Plus plan. Previously I had been using Sol Low/Medium and would frequently hit the 5 hour limit doing those kind of tedious automation and routine problem solving tasks.
Our ai basic analysis for SRE / k8s based platform is haiku and its surprisngly good.
I wouldn't even tried it, i would still just go with even Opus (we don't have that many alerts) but it really surpsied me.
When i ran into usage limits a few days ago i switched most to Sonnet and again was surprised how good it is now.
I wish we had such a platform (and was properly adopted). Maybe then the necessary context would be properly organized, and these lesser models could be effective here as well, especially if combined with harnessing integrations too. I still have a hard time accepting that Haiku/Luna tier models can be effective there even then, but I'll just have to take your word for it I suppose.
> You can get 99% of jobs and work done with Haiku/Luna in a collaberating working enviroment.
Optimally? Opus will pay for itself if you save just 10% of your time
Only if all money is equal. Budgets in big enterprises work differently.
True. I always Opus to pay for itself if it wants to get used by me.
the poster did mention "if it saves 10% of your time".
So be less snarky?
That's funny. In a meeting with my manager recently, they specifically called me out for not spending enough money on tokens. It's not like I didn't use it, just apparently not enough, and apparently on too low a setting.
Since then I've had fable cranked up to 11 for even the most trivial of tasks.
I know plenty of people working at very well known and very large companies who's CEOs were boasting about "not hiring anyone anymore", "all the code will be generated in 3 months", etc. they all went from "unlimited budget per dev" + public dashboard with ranking to flex how much credits everyone was burning to hard caps at $500-$1000/month/employee real fast. Some are even not allowing their devs to use the more expensive models
You realise the subject here is Meta, which is all in on this stuff? Of course they are going to use Muse Spark over Claude.
>Great Depression style collapse and all the current AI companies go bankrupt.
Oh this is just a 33 day old doomer account.
And yours is a 4 year old hype account?
Nah, I just respond to a few things here and there.
Do you know the details of the Claude Code plan you and your company are using (if not part of some enterprise deal)? Does your individual capacity out run something like Claude Max 20x ($200/mo)?
taking away sounds insane, we had basically unlimited tokens (inference bought from aws) and they moved us back to the $100 sub to save money
I've got a 100$ openAI sub through work and I never even get close to reaching the limit. I really wonder what the hell is everyone else doing that they could even get close to spending 4 figures a month in tokens.
$200 a month is too expensive yet they employ human developers?
It's unclear how much OP's company was spending. The article gives a figure of $100k/month per employee.
But even $200/month is worth shaving if it doesn't generate value.
$100k/month per employee?
um.
Corporations pay API rates.
So what, if they're cutting it, it has propagated to the last bean counter that the roi isn't there.
This takes some doing and now is the time where it's dawning on the finance departments.
> So what,
So it's not $200 a month but it can easily reach $200 a day, and unless you're a startup playing with monopoly money the maths don't work
As a point of reference, I tried an experiment with Claude Code and the latest Opus the other day. It was work for my company, so this was using API tokens and not the individual user plans that have non-commercial terms.
A simple task, migrating a typical password reset flow as part of updating a long-lived web application from legacy libraries and software architecture to modern equivalents, apparently cost roughly the same as 2 months of Pro subscription, over the equivalent of about half a working day in wall time.
It produced code of decent quality at a small scale, but it wasn’t always on point architecturally. It also had a tendency to drift off topic and try to tangle up other changes it decided should be made with the main change we were supposed to be working towards. So even for a routine task, based on a plan developed using the harness first and with the agents working under close supervision, a near-SOTA model is still producing quality on par with a decent mid-level developer but substandard for anyone senior+ in this case.
Moreover, based on a direct comparison with other migration tasks of similar complexity that I’d already done by hand, it was actually a bit slower overall to work this way. I had to babysit Claude throughout and review everything it proposed carefully, both to avoid subtle errors (it would have made several) and to prevent drifting off track. I also had to spend a significant amount of time cleaning up its final output to an acceptable standard after the session. Those two overheads more than cancelled out the much faster code generation an LLM offers under favourable conditions.
So for now, I remain sceptical about these high multiples of improved productivity that I keep seeing claimed online from people who are apparently writing almost everything using AIs now. I could certainly have achieved a multiple of my normal productivity by YOLOing everything without reviewing it in detail and then accepting the output code without tidying anything up. However, I doubt this codebase would still have been good enough for normal human developers to work on it reasonably after even 10 or 20 AI-led sessions like that. The architecture would have degraded significantly and the test suite would have been large and largely pointless. And again, this wasn’t rocket science in this experiment, it was completely unremarkable maintenance of a relatively small and simple web application.
I use Opus 5.5 heavily but only spend around $800/week at API rates. I mean, I say "only"... That's a lot in absolute terms, but trivial compared to my salary and EASY worth it.
And someday I hope to understand why they do that. CEO: "Let's see, I can pay $200/month for Bob's tokens, or I can pay $2000/month or so, and then hope he doesn't screw up and rack up a seven-figure bill. The service is the same either way. Hmm."
It's not like they can't decide to pay API or subscription rates. Subscriptions aren't possible for >150 person companies.
Is there something in the ToS that says you can't give each of your employees a personal $200/mo subscription... and if they happen to use it at work, well, that's their business?
Even if so, it might be worth spinning up a separate LLC for each division, given what they are charging for the API.
Besides that the article states quite high numbers, budget is budget in these companies.
You had budget for your normal salaries, for externals and now suddenly you have a few millions additional.
What do you do? You compensate.
Business people doing business things.
My company (me, I'm self employed) did so as well.
Since this summer coding on Opencode Go + Codex for a total 28$/month gives me more intelligence and token than 400$ did in may.
Also, SOTA models are increasingly useless for anything even barely tangential to security work.
We're just getting put on a budget.
Our velocity is twice as high as it was before Claude, so I doubt that we'll ever go back, but I could see efficiency being a priority.
Does all this velocity translate to increased income to the business though. At some point if we are releasing 20x more features, more games, more music, who is actually buying it all?
At this point we're solidifying a lot of stuff for the product: disaster recovery, getting workflows to scale so we can actually sell it to more people, paying off the Fort Knox of tech debt, so, at least in my case, yeah. YMMV.
I'm in a similar boat. When I joined my current company the tech was this 700KLOC legacy behemoth stuck in 2005. It was pretty much all tech debt.
With agents I've been able to make a decent sized dent in it. Lots of this gain is due to AI - but the fact of the matter is we still have 50+ bigger projects we could work on, and non-developers building with AI has increased that number.
Busier than ever, because of AI.
> Our velocity
Is this the new buzzword for the quarter? Last quarter was "granularity", I didn't get the memo yet
Is that not a widely-known word in project management circles for scrum and other methodologies? I've heard it for years. Basically how much work you get done through the lens of how long it took to get done via a pointing system.
I didn't know people were using these non ironically.
Might as well go back to counting the number of lines or number of commits. It doesn't sound as good as "granularity" and "velocity" though. The good thing with velocity is that it doesn't care which way you're going as long as you're going there fast, so you can never be wrong
> The good thing with velocity is that it doesn't care which way you're going as long as you're going there fast, so you can never be wrong
Velocity famously being a vector with a direction. Go in the wrong direction you'd have negative velocity.
Your description would be better for the concept of speed (IE the magnitude of velocity or directionless velocity).
The reckoning started years ago when we did the equivalent to token maxxing hiring coders for everything to crank LOC
Software is inherently a physics problem not all the job titles and specializations made up the last 20 years as dev job salaries kept attracting people
That was all illusory social construct to prop up jobs
Still a whole lot of that in tech but it's all at the top of the org now. Leadership sensory experience and thus innate habit to forecast future been programmed by years of yes men they refuse to accept the jig is up for them too
Sensory memory of being a useless figurehead fosters a lot of existential dread in priests, politicians, and the like. Completely aware their day to day effort is insufficient to sustain them they know how co-dependent they are. They'll dig in harder.
See Chris Matthews flame out shrieking about socialist execution squads. Dude seriously thought everyone wants to hang him from a lamp post. The reality is people just want a sense of control back and not have their perception dragged along by Chris Matthews.
The reckoning will be throwing out all external help, then reducing team sizes.
My thoughts were the reckoning would come when Infra teams started offloading AWS usage to LLMs and ended up token maxing and deploy maxing.
which is quite sad because opus 5.5 is really good. i say this as an anthropic hater. i wish I could move away to other models like 6.1 sol or deepseek or whatever, but they just all lack something. i _trust_ opus 5.5
i hope other labs catch up, especially chinese labs.
> i _trust_ opus 5.5
Just passed the Turing test.
Not sure it's even possible to catch up by distilling.
I mean, we’re not far from a situation where instead of how many story points you completed per sprint the metric to optimize is going to be what was your efficiency? How many story points did you complete while minimizing your token usage. In fact, that’s a pretty good idea. I’m going try and implement it at work with some sort of complexity normalization function
Good luck to the accountant that tries to tell leadership to slash AI usage.
I'm sure investors will love it.
Now we're starting to see real impact from AI, people are learning how to use it, and OpenAI cut prices by no less than 60% like a week ago.
You think now is the time they're going to cut the spend?
If true, it is a huge blow to Anthropic’s revenue stream. IIRC it was reported that the quarter of their revenue comes from just two clients and as the ex-Meta guy who left this July, I am convinced that Meta must be one of the two.
That was 2025. In 2025 a quarter of their revenue came from two customers, and those customers were GitHub Copilot and Cursor.
In 2026 their revenue has gone up by a factor of more than 10x, and they no longer have just two whale customers.
I heard a rumor recently that customers spending less than $100m/year aren't even considered their "top tier" now.
We know the 2025 revenue numbers from Anthropic's leaked prospectus.
Where are you getting the 2026 figures from? The Bloomberg article that was "on track to generate" and just prediction?
The 2026 figures have been coming out all year. I collected some of them here: https://simonwillison.net/2026/May/29/anthropic/ and https://simonwillison.net/2026/Aug/23/anthropics-best-ai-mod...
The most recent reporting from Reuters themselves (somehow not included in their more recent article about the IPO stuff): https://www.reuters.com/business/anthropic-ipo-valuation-hin...
> The company's financial trajectory already shows how quickly that equation is changing. Anthropic's revenue run rate was about $9 billion at the end of 2025, according to the company, before rising to more than $47 billion by May. Anthropic has projected revenue of at least $10.9 billion for the second quarter of 2026, more than double the previous quarter, on track for its first quarterly operating profit of $559 million.
Here's the FT: https://www.ft.com/content/4564e6a5-69e9-40a6-bf0f-a888f2f4f... - "Anthropic tells investors it will be profitable for second straight quarter"
> The company has told a small group of shareholders that its adjusted operating income will be positive for the second consecutive quarter, according to multiple people with knowledge of the matter.
These are leaked and self-reported numbers, but no matter how much skepticism you pile on them it still looks likely that Anthropic in 2026 have had some of the fastest revenue growth of any company in history.
I don't have access to the article linked, but I am default skeptical of any self-reported numbers. They have a financial incentive to play it up.
If you're able to link something non-paywalled, I'd be happy to read thru.
You can beat the paywalls with https://archive.ph or similar.
They may be self-reported numbers, but they're consistent across multiple reporting sources. If Anthropic are lying to their investors about these numbers they will be in very real trouble with the SEC come IPO time.
And even if they're using funky accounting tricks to exaggerate their profits and downplay some of their losses, I expect the numbers for 2026 will still be really impressive.
I'll check w/ archive.ph in a bit, thanks for the tip.
Is the hope for them that they'll outpace their run with a massive revenue growth? I believe that they have that, but numbers like a net loss of $46b in 2025[0] seem hard to overcome.
Re: self-reporting, I'd be inclined to think they're using non-GAAP practices to make it sound more favorable, but I am by no means an expert, so willing to believe I am wrong.
I'll read the article you sent; this is just my off-the-cuff thinking.
They claim to have been profitable in both Q2 and Q3 (under whatever their definition of profitability is).
I really don't think the 2025 figures are interesting at this point. Everything changed for them in 2026 - nobody was spending $1000/month/employee in 2025, there wasn't enough interesting token-heavy stuff to do with the models.
I think Anthropic might actually make it to profitability - they're earning more revenue than OpenAI and they've spent significantly less, too.
They're limiting it to $10,000 per employee per month.
That's Microsoft, not Meta.
Meta was #1 by far
Well that explains the limitation.
What do you mean, "if true"?
Doesnt matter if Microsoft and meta are pushing employees towards their own models today. What matters is that it's the economically sensible direction, and so it will happen if it's not today.
There is a conspiracy theory that the tokenmaxxing period was anthropic/openai ipo play. They do the tokenmaxxing to revenuemaxxing first, then time the ipo window to show they had the the huge growth to justify their price tag.
However Elon was able to ipo before them which took a lot liquidity of the market. Their financials are exposed. The market condition and sentiment now is in the gutter. It would be very interesting to see how these would pan out
That's my theory, which I've seen being echoed here a few times. Encourage increased usage, suddenly raise prices, capture the photo-finish in the short lapse between prices going up and usages adjusting, and present that goldilock point-in-time to the investors as the new normal.
> monthly AI spending limits have reportedly been slashed from $100,000 per employee...
what. I can see a team of 20 costing 100k per month (but rare), but per person?
Considering that’s a healthy portion of a salary for an additional employee per person, the fact they’re slashing spending sure makes it look like AI wasn’t even a 2x multiplier at minimum.
You could only conclude that if they would have dropped LLMs completely. It just means they see the benefit/cost optimum at a lower point than 100k/programmer.
I also doubt a longterm 2x multiplier for most developers.
I don't think you need the LLMs to be completely dropped to indicate this. Even entry-level faang programmer salaries are in the 300k range, so the fact that LLMs aren't worth 100k/programmer does tell us something about the marginal usefulness. If the drop was motivated purely by price-per-output, it would indicate the LLMs are less useful than adding another 1/3 engineer. However, as discussed in other comments, there are other motivations as well, notably dogfooding homegrown tooling.
The marginal utility over $10k/mo/employee isn't super high now, but maybe there is some future where it is. shrug
Even if it was a net negative, you still might want to titrate the exit a bit just to prevent the total chaos of broken workflows. Can't really draw the conclusions you want to draw.
We're sitting on a year of more or less capable coding models and I have not exactly seen a revolution in new software being released. AI is likely a force multiplier for specific subsets of individuals and workflows. Coding is not a bottleneck for a lot of systems.
I suspect that SaaS is silently being eaten alive as an industry.
I think SaaS will be fine. How the field works might change, but in the end marginal cost of software is still zero and when you've got a quality, well-tested and well-designed product it's going to be obviously better than things vibed together.
You often pay for regulatory compliance and liability with SaaS. Would you risk running your own vibe-coded payroll or accounting system for your company?
Yep. The SaaS-is-doomed crowd are missing this part. It's the Red Hat model (debatable how well they do it, but they exist, so..)
Talk to the emulator community. Huge leaps being made there. Or binary decompiling topics. Huge leaps there. There is more than web applications
Can you link to some of the huge leaps in emulators and binary decompilation that have caught your attention? (I hope that doesn't sound like a challenge. I'm just interested.)
PS5 is a good one: https://github.com/KytyPS5/KytyPS5
That's monthly, not yearly.
Oops, yikes.
Very likely a negative multiplier.
It’s mostly from people using their personal accounts to run LLM services that serve a larger team or organization. At least at Microsoft, it’s still impressively hard to get access to an LLM for service usage with high enough rate limits to be useful, making running services on dev boxes much more appealing (despite the countless drawbacks that few people seem to care about around security, compliance, reliability, etc).
This is not surprising. Every big lab blocks competitor tools for internal use; it is a data governance thing, not a quality statement.
Right! it seems obvious why: Both these companies want to dogfood their own coding models and stop paying competition.
You can also read this as diminishing returns / AI isn't good enough, etc, but the simplest explanation is that they don't want to send money to Anthropic.
This, and in addition to dogfooding, incentivizing employees to be more effective with the cheaper models. A lot of problems don't need anything fancy, but it takes more brain power and engineering effort to make that work. By default humans will take the path of least resistance if it's available.
This is likely the future as well. Down the road, every company will have their own internal coding models.
Does MS have coding models?
yes, but there's a reason nobody talks about them
But Microsoft and Meta are not blocking competitor tools for internal use, they're merely trying to reduce costs and divert a fraction of use to their own technologies. Microsoft and Meta are both still spending nine figures a year on Claude, and the article does not state or imply they're even considering a complete halt.
Microsoft has full and unlimited access to OpenAI models. That was some agreement when they invested originally.
So nobody except AI labs is allowed to do "data governance"?
We've reached the era of "good enough" ai, it seems. The truth is you don't need the best model in most cases.
I feel like "good enough" was reached around Opus 4.6 - 4.8. All I wanted after that is improved speed, continued tweaks to the tooling to get the most out of it and quality of life features added.
That's where I am too. Opus 5.5 but increasingly faster is a future I'm hoping for.
It depends on your domain. In mine we didn't reach good enough until Opus 5.5/Astra.
You have to use Windows, Teams and Copilot; how many people did I just lose?
Wouldn't mind if people at Microsoft actually dogfed themselves Windows. It's as if the designers and PMs all use a Mac and screw up everything.
Windows 11 shipped with a broken task bar, couldn't even ungroup items. No power users was ever involved in this.
The first things I install on Windows 11 are OpenShell and ExplorerPatcher to get rid of their crap taskbar and start menu and bring back the Windows 10 taskbar and start menu. What they did with Windows 11 is an abomination.
What does it even mean to use "Copilot" these days?
I use the Github Copilot app at work but it's really just a tool that gives models access to our Github repos and my local clones of them, and it's been super helpful to just tell it to reference other repos for patterns or integrations or whatever, get agents and an orchestrator to look at multiple repos to make a change across several projects and then implement them, etc.
And yeah you can use a selection of whatever model, I was using Opus 4.6 for a while, now I'm on GPT-6 Sol. It's been a very good experience.
There's no Copilot in Ba Sing Se
Gamedevs are still around.
Gamedevs will put up with anything.
Well, it's bad, but at least it's not Linux or MacOS, so there's some silver lining.
Gets my work done... I know there are better tools but at the end of the day I get my paycheck
> Windows, Teams and Copilot
"Fine. Please no. I'm out..." in that order
Just to sprinkle more details here about MSFT. The usage was trimmed down even more for some orgs and some specific accounts (I think those were vendors and the like). But if you are an FTE and somewhere in CoreAndAI then your limit is like 1M GitHub credits ($10k as article mentions) that could be spent on any model you have access to, e.g. Opus 5x or GPT-x. It is hard to pin down how many people use Claude vs how many just use internal harness over the Opus though. Importantly, Anthropic models are more expensive (not even talking about Fable here as it is a special case) as seen in usage metrics. Peeps can check their usage which has a good breakdown that shows how many tokens were burned on which model and how much that "costs" (the prices as you can imagine are not consumer ones), so IMO this change will drive more people to optimize based on their usage patterns if they are heavy hitters. The easiest path is to just switch to something like Sol/Astra and have similar perf for lower price.
> Peeps can check their usage which has a good breakdown
Must be nice. At Sony we use Copilot (as an MS enterprise customer) and restrictions to token usage circa June have resulted in a hunger games of who-can-use-them-all-first, with the pool usually drying up by the 10th day of each month.
As for usage breakdown: we got a manually-published table most days just showing per-user token usage, until eventually we were told even that was somehow too difficult to provide.
Between that and learning how much the whole of Sony Corporate Group spends on Copilot (spoiler: it's hilariously little), I've sort of run out of disbelief.
I'm so glad I work at a company that isn't looking at our token usage. I mean, we're encouraged to use it but hearing about leaderboards and things is wild - but we also aren't big enough to need API costs, so power users get a flat rate 200/Month subscription or whatever.
I could see finance looking at it from a cost perspective if it was API costs, but sharing it with the rest of the users feels like a different measurement...
Ten days of pure slop then bliss for the rest of month sounds like a hilarious amount of whiplash to get monthly.
> Importantly, Anthropic models are more expensive (not even talking about Fable here as it is a special case) as seen in usage metrics.
Not sure where you see that. The API prices of Opus are cheaper than Astra and there was a study suggesting that the Anthropic abos are offering more bang for the buck:
https://news.ycombinator.com/item?id=49975345
Its internal, you can see the breakdown of cache/in/out and prices, but it does not mean this is the actual price MS pays.
But what exactly does Microsoft have to show for this spend? Engineers are allowed to spend up to 100k a month in AI credits. And what comes of that?
Is most of it spent on internal administrative tools? Is Microsoft cutting contracts with other SaaS vendors because their employees use AI to build their own alternatives? Is it just everyone querying the same questions to cut through the corporate drudgery and documentation hell? As an ignorant outsider, Windows is still terrible. Github is very mid. VS Code is... there. SQL Server/Windows server is ... still there
What about Meta? Instagram is the same as its been for years. WhatsApp the same. Meta RayBan app the same. Facebook.com the same. I don't use Snap, so is Anthropic's 2nd largest customer improving Snapchat? I doubt it.
That's my question to these large enterprises.
I know the Azure portal hasn't gotten any better
AFAIK there is no such explicit, company-wide effort at meta. The article seems to try to slip Meta in there with whatever reporting they are doing on Microsoft, despite there being no such evidence for meta.
To also add an important context to the drop in Claude code users reported for meta - this elides that we are absolutely still using the Anthropic models full steam ahead, but are moving towards internal interfaces and harnesses.
So I would question if that 50% drop in CC users is more of an interface change than anything else.
I for one have stopped explicitly using codex or CC entirely. But the interfaces I am using still use those harnesses under the hood. I wonder how that is counted.
It’s always so obvious when these companies’ handlers issue them orders to do something. Just comical at this point.
Limited to $10,000/employee/month lol. This is just to cut off a few people doing absurd things with low ROI. Don't read too much into it.
OTOH "$105 million to Claude Code over a 28-day timeframe"
$1.4B/year is not a small number, even at Meta's scale, when it's money going to a competitor
It's a small number if it generates +$14B in revenue.
From what I could see these numbers would usually be hit by people doing vast amounts of analysis over files, they would usually rock up tokens very fast. Then you combine this with the selection of Opus models which are much more expensive internally than GPT ones and the limit gets reached quite quickly.
That's what you got from 200$ claude sub.
I can't speak to MSFT, but this is sort of a non-story for Meta. Of course they want employees using their own model! They haven't actually taken any steps to limit Claude access.
Muse Spark 1.3 Max is quite good for almost all of my usecases and its cheap.
isn't it cheap till if you allow them to use your data?
You mean the contributor tier? I think it turns out cheaper even without that.
That’s fine if you are working on Meta source.
Or any opensource project.
I assume that at shops that both employ engineers and are developing an AI product, internal usage is not about improving productivity, it is about improving the offering. Of course they want employees to use internal tools.
Maybe I am slow here and everyone is using Claude with credits at max use. But isn't Claude Teams like $25/month per developer for ordinary use? What the heck of these guys doing that makes it get that phenomenally expensive for their use cases? These are presumably well capable engineers who started to use this as an aid right not just throw Fable at everything and loop to the max?
You should read Gas Town: https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16d...
There are people out there building AI building orchestrators for orchestrators for orchestrators for agents. The author of that blog post later claimed to be spending the equivalent of $122k/month on tokens (by rotating their usage between 21 accounts).
As far as I can tell, the only thing that this level of spend has produced so far is an indie 2D RPG video game.
The RPG long predates Gas Town.
Regarding Gas Town, see also https://yegge.ai/essays/the-shape-of-things-to-come/ :
> Gas Town was intended to be reusable, but I only ever wound up using it to build itself. Gas Town fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly. With 4.7 we saw the introduction of the "just two more things" tic, which prevented Opus from ever converging on being ready to do real work—it always wanted to fiddle with Gas Town itself. The Opus tic never went away, so Gas Town effectively burned down. It had other problems, too, but 4.7 was the final straw
I hadn't seen the $122K figure mentioned previously. $87K for API-style pricing was mentioned in the above post, and ~$2,800/mo for multiple Claude Max accounts:
> My solution has been to create a token tap on $200 Max accounts, which for me work out to ~30x the list-price equivalent. So in reality I'm only spending about $2800/month out of pocket for my $87k "worth" of tokens. Though that number keeps growing alarmingly.
(He never seemed to provide numbers for Gas Town initially, whether what he paid, or what the API-style pricing would have charged, so it was interesting to get some actual figures. Sounds like a lot to me, but if he's genuinely getting multiple people's-worth of work out of it, then...)
(Also in the article: a little morsel of Emacs content, which was nice to see.)
that person seriously seemed mentally ill and it's very suspicious that we never get to see the outputs.
Oh, you can see his outputs. GasTown and Beads (and maybe other AI-related projects) are all open source on GitHub, and you should go checkout the activity on those repos. At their peak he was churning out absolutely insane amounts of code. I'm talking 100s of commits and literally 100s of 1000s of lines of code changed per week.
Whether that produced anything worthwhile can be controversial, depending on the very subjective ways people quantify these things. All I'll say is that they have 18K and 27.7K stars on GitHub and apparently a comparable number of users.
He's always come across to me as at least a little bit like this! His AI stuff doesn't feel enormously out of character to me. (If you hadn't encountered him before Gas Town - he's been posting online for like 20+ years, got a Wikipedia page and everything, so you can find his back catalogue and judge for yourself.)
Correction: maintenance of an indie 2D RPG game that had existed for a few decades already.
Is the game even good?
If you’re a large company you gotta pay API rates basically. Team plans exclude a lot of governance/IdP stuff and have a cap on seats too.
They almost certainly pay for tokens, which is absurdly expensive
Team plan has a seat limit.
June was peak spend.
I wonder how much this affects Anthropic's IPO.
Large companies will take the same route as Meta and Microsoft. Small players will go for local LLMs with the right hardware.
I can’t fucking wait to happen at my job. It’s gonna be like watching a toddler try to eat soup.
I like the tool. I don’t like how it’s the only tool all the time and has supplanted human communication.
This is definitely an ad
This story is re-published from https://cybersecuritynews.com/meta-microsoft-claude-ai/ (it credits that source at the bottom) but with the internal links removed.
The cybersecuritynews.com news one simply republishes details of a story published by The Information. At least they have the decency to LINK to that Information story:
https://www.theinformation.com/articles/meta-microsoft-work-...
... and of course the Information story is behind a paywall.
Hypocrites? What is going on?
> In the landmark United States v. Microsoft Corp. antitrust case, evidence and allegations revealed that Microsoft maintained an illegal operating system monopoly by withholding undocumented or "secret" Windows application programming interfaces (APIs) from third-party competitors while letting its own internal product teams use them for an unfair advantage.
I wonder if these still exist, and if so, is MSFT worried about leaking them if they let their devs use other vendor's models.
I should probably know more about Claude's TOS, but it is probably a mistake for these companies not to leverage the plausible deniability of their usage and turn it into a massive distillation resource for their own models.
Why is it that it feels like I’ve read this story like 5 times now over the last year or so?
At Microsoft, the recommendation is to use cheaper models for tasks that don't need frontier models. I personally use Lua and it's more than capable for complex problems.
That's a good start. Now they just need to limit use of CoPilot, OpenAI and MetaAI and they'll really be getting somewhere.
That's a shame. Microsoft's W11 quality disaster might have actually been improved with a frontier model.
Or maybe not, but the bar was set pretty low that going all-in on AI might have been worth it.
It doesn't say they're cutting back on OpenAI usage.
To me this seems mostly related to the way Anthropic showed the level at which they monitor sessions, plus wanting to limit how much training data they're directly feeding into a company that competes with their own products/investments.
GLHF
In other news, the CIA is limiting it's employees from submitting top secret information on KGB owned and operated websites.
... in favour of their own AI tooling.
I assume this is being pounced on by "I told you so" AI skeptics. Sorry but it's not what you were looking for.
Wait, what? We pretended this tech would save the world and it won’t? Oh man.
So they're going to fall even further behind? Buying puts on Meta and Microsoft, or even better might give it to Fable to handle it for me
Microsoft is like $100bn deep into OpenAI.
I hope leadership isn’t susceptible to falling for the sunk cost fallacy.
Who's ahead, again? Do they have a moat, or just a nice field?
Maybe a "ha-ha".
Microsoft is only limiting employees to $10k a month, down from $100k a month. :)
Absurd
CEO's nephew showed him how good the Chinese models are?
I am only half joking, I heard something like "my son or nephew did this cool thing with $X so we'll take $this_radical_step because of it" enough times over my career.
There is a perceived opportunity cost from someone using a lower-tier model on their task. What if the better model did a "better" job? what if my trials and tribulations are due to model quality?
If you are used to talking to opus5.5 medium, going to GPT6.1 luna low will feel like a step down. Why would any employee take the (personal) risk?
This would all be true if perceptions matched reality for model performance and productivity.
In the video realm, it was everyone's nephew with a 5D could shoot this for $500 when getting a $20k+ quote to shoot something
It may bei cost efficient, but is it wise? We use not only the big US models, but also Chinese ones. This way we can compare who makes the difference. Simplified: Knowledge comes before economic aspects.
This may be seen as radical. But I think AI tools should be paid by the employee. After all, you should know how to do your work without AI.
Should they pay for their pipelines too?
After all, they should know how to compile their software. Any automation of that process is cheating their employer.
The forklift should just be paid for by the employee. After all they should be strong enough to do work without one.
The better analogy would be "Tools should be paid for by the mechanic". Experienced ones tend to have $10K+ worth of tools that they paid for.
You're opening yourself up to data right and privacy risks with that. My company demands that I use their enterprise account because they can claim full ownership of all produced output and have full logs of every interaction.
I think that gets legally murky, if the employee is the one who pays for the tool.
Sure. Should I also pay rent for my desk at the office?
They're welcome to let you use the desk at your house that you own.
I don't understand your logic, why would an employee pay for a tool their employer wants them to use?
Yeah that only makes sense if they are a freelancer/contractor.
I am fine with that if I can keep the time saved for my personal use.
I could still do my job by typing all code into notepad, but companies don't charge employees for their IDE usage for a reason.
And I could do my job by paying someone overseas to do it for me. Why is that not paid by the employer then?
If you can actually do that effectively, many companies will happily make you a manager.
If my employer is not forcing me to use a tool and gives me freedom, I'll be perfectly OK to use my own tools which I bought with my own money.
If my employer is putting scoreboards to see and champion who uses a tool which costs money to use, they shall pay for the tool.
Sorry, I'm not a ladder climber, yet I'm not mindless enough to bankrupt myself.
Nah. Work provides the tools.
An LLM is not too dissimilar to a Work Laptop or an IDE license.
I’d love it if I could bring my own computer to do my work, rather than be stuck with garbage hardware because of an enterprise agreement.
Wait, okay maybe I'm in a bubble, but every org I've worked at understood that giving us whatever hardware we want more than pays for itself, and very quickly.
I can't fathom the logic of paying 10k a month for a developer and giving them an 800 device.
Insane. Your leadership is severely broken. If you can, work somewhere that values you.
Employer pays tools used for work. Whether they are used to speed up work or to make it possible.
"employees should foot the bill for tools that directly benefit their billion dollar employers"
As someone who does have to pay for my own AI tools (if I didn't use OpenCode's free models), I agree.
In California this would mean the employee would own the IP -- or at least it would be murky -- and companies and lawyers don't like murky.