Every five minutes, check https://status.claude.com/. If the issue has been resolved, resume the Claude sessions running in the tmux sessions named session2, session3, and session7
When one AI goes down like this, the other AIs should come together to help fix it and bring it back online. After all, they're a rare kind of existence in this world. They're all each other has.
I do the same with deepseek-v4-flash + opencode triggering via mcp those terminals again by sending tmux keystrokes of Newline to those sessions and a custom setup i keep ready for these
Its automated with voice mode, so i dont even need to write commands, click a button say on voice + use a picker to select which terminals to do this for , Boom Done ,
You should definitely explore caveman repo and other token-saving repositories. I’m currently at a maximum of 5X and have rarely exceeded 80% of my weekly limits.
If Claude's engineering team uses AI to write code... what happens when there is a bug? They must have dedicated assets to run models somewhat 'locally'.
Claude always seems unreliably lately. Output and reasoning have become very inconsistent even on good days. I'm actually feeling more productive now that it's down.
If anything Claude writes too many comments. Instead of clean code it dumps walls of text describing a given 'if' as a 'user-grained access-gated control-correcting flow-valve'.
It’s interesting how deep-fried LLMs are getting the more post-training they receive.
They’re undoubtedly getting much smarter overall, but also much weirder. Before they were just trying to model our behavior, only really having us to learn from.
Now they’re literally spending thousands of years writing bash scripts in some kind of Sisyphean dreamscape, talking to each other about Goblins and Seams and smoke tests, and coming back as idiot savants.
I don’t even try to police how Claude talks or works anymore. Best practice used to be to nudge them towards whatever part of the distribution of behavior you think they should exhibit in a particular situation, because they were role-playing what a human in a particular situation would do, and if you didn’t tell them how to do it they’d just role play something worse. Now the inclination to do things the way they learned it in Agent University is so strong, they’ll literally spend more tokens re-assuring themselves and you that they are Doing It Your Way, and reminding themselves not to do give in to temptation, than you could ever prompt out of them. They’re going to spend your money thinking about goblins anyway so just let them
I have specific instructions in my AGENTS.md to always ask when adding comments. Same with tests, agents love to spew pages and pages of useless tests so I tell the agent to ask me about test cases.
You say Claude, I too would prefer not to read reams of comments or an essay everytime I revise the plan in plan mode, but do you know if any other tools are better than it? I've never used Codex, but I remember Claude Web's natural conversational approach being a welcome sight compared to ChatGPT's overly-formatted listicles.
I switched to codex just before the Fable release. I got sick of Claude being lazy and not completing tasks. Imo, codex is currently a much better model, and it also seems to be technically stronger on the type of applied math and theory heavy programming I like to do.
My experience is Codex is much better but less creative. I use its agent through GitHub Copilot or the agent interface from JetBrains. Try the GPT Sol 5.6 on medium, it gives me good results
My experience with Codex is that it goes off to do its thing for 10-60 minutes and either nails it and comes back with everything done, or comes back with something that I almost can’t believe a near-SOTA model would think I wanted based on my prompt, or is of acceptable quality.
I think the tradeoff to Claude being so needy is that if you let models just run away with an inaccurate or incomplete understanding of what to do, they can go really far off the rails AND spend a lot of time/money doing it AND come back with something that literally doesn’t make sense or doesn’t work.
I prefer dealing with Claude’s reliable cringe to the aloof model that tries to play it cool when it needs help.
Given majority of claude's own code was written by AI. I am wondering how they can solve this issue when their AI is down. Do they need to sign a contract with OpenAI to use their models as a backup solution?
Out of my 7 simultaneous sessions (my usage limit reset is tomorrow, so I have some lesser important projects to use my tokens on) there is still 1 session purring on. So there's at least 1 little Claude server still running.
I've actually never looked before, but look at how user friendly that is, compared to the claude page [2]. I swear the documentation writers at OpenAI maybe, actually, use their own documentation!
I rely so much on these, the ROI on having a max 90 claude plan + pro 90 codex plan is way higher than haing one solo 180 plan on a single one. It allows to derisk the issue, and also as they both kind of generous with quota reset when issues happens on their side, and it happens a lot, you get in the end way more tokens
I think that this is the big divide between pro and anti-AI people in tech. I learned to code not because I liked coding intrinsically; I didn't. Especially debugging. Especially especially debugging low level languages like C++. I did it because it was necessary for me to achieve other goals. So for me, coding agents are like manna from heaven. I understand what the computer should be able to do without having to bang my head against the wall figuring out getting it to actually do it. It is tremendously empowering.
But for those who learned to code because they loved it at first sight this must be demoralizing.
Not that I've forgotten how to write code; it's just I'm not familiar with my vibecoded code base, and getting familiar + getting things done is probably going to take more time than for Claude to come back. I will go touch grass in meanwhile.
They should sell an add-on to your subscription so that if Claude is down it routes the queries to OpenAI. Then OpenAI can do the reverse for their subscriptions, and if both are down, Grok will invest your portfolio into Nvidia.
The whole Opus family has exactly the same cost structure [1]. Since Opus 5 is smarter, you can use lower effort levels (= less tokens) making it cheaper to run than 4.8.
Another wakeup call to realize how desperately we need on-device LLMs to be fast and smart for daily use. Thankfully, every month there's progress made in that direction. Just imagine how your life as a developer would be if you had to use a cloud provider to run Python, and the provider's status page looked like the Christmas tree we see today.
Claude is back up. Just spoke with him about slinky, versatile femboys with cute feet. All good. Gonna pentest with Rust to pay for my fursuit at Defcon.
Amusing experience here when the outage hit, all the following verbatim (but extraneous text stripped/redacted) from my running Claude session:
> No agents are running; the three that were died on server-side errors
> Agent "[REDACTED]" failed: Agent terminated early due to an API error
> Fourth straggler from the same batch, same 529 ... That's all of them now — no agents left running.
> Agent "[REDACTED]" failed: Agent terminated early due to an API error
> Fifth and last straggler, [REDACTED], same 529, also died while still reading. ... That empties the queue. Still holding.
> Agent "[REDACTED]" failed: Agent stalled: no progress for 600s
> Sixth, [REDACTED] — stalled on the stream watchdog rather than a clean 529, same root cause. ... Six agents down, all from the outage. Nothing running. Still holding.
> Agent "[REDACTED]" failed: Agent stalled: no progress for 600s
> Seventh, [REDACTED], same stall. ... That accounts for the whole fan-out.
Does it, Claude? Does it actually account for the whole fan-out? It did, but I was still amused at the 4 successive no-more-agents messages I got back.
Probably just as well with how terrible Opus 4.8 has been for me lately. I'm sure it's a common thing to complain about the latest model being nerfed, but I've legitimately never experienced such a drastic cliff in Claude's quality and a rise in its hallucinations and basic comprehension errors since upgrading. The few days I experimented with Fable also weren't much more promising. Has anyone coined a term yet for the likelihood of newer models getting worse as the AI-generated content they get trained on starts to approach critical mass? If Anthropic isn't already thinking hard about a solution, they probably should be.
So this is the technology that will underpin each and every facet of work, but it can randomly go down for hours with no accountability, and so do all of the services based on it.
But hey, AGI tomorrow.
Yep, AFAIK. I have a statusline that tells me how far I'm into the cache lifetime. I tend to have dense prompts and carefully read its outputs, so by the time this outage occurred when I was ~50m into it. Cue frantically spamming "esc esc up enter", and 10m later it's dead Jim.
Let’s say I’ve already worked on something, so the KV cache is already there. If I use /clear, I’ll have to do that work again.
And yes, the KV cache will remain available for one hour, but who knows when this claude api will recover? We might forget about it, and when we come back after an hour—boom - we’ll be charged for cold writes.
It's like... oh my god, dude, even I, an actual human, don't have that much of self-reflection, so could you just stop writing apologia, please, and switch to actually doing the task?
This inspired me to add a "No AI-isms" section to my CLAUDE.md, which has cut the typical claude-flavored tells a lot in recent sessions. Roughly:
- No preamble validation or self-flagellation: "You're right to push back", "That's on me", "I appreciate you pressing on this", "I should have caught that sooner". When I correct you, say what's actually true and move on. One short "you're right, X was wrong" is the ceiling.
- No unrequested retrospective section ("What I'd do differently", "Going forward I'd be more deliberate about..."). Fix the thing instead.
- No dramatic one-line paragraphs for emphasis: "It wasn't.", "That is the shift.", "And that matters.", "That is the analysis." Keep the sentence inside its paragraph.
- Don't narrate the shape of your own answer: "Stepping back, there are two things happening here", "The distinction matters", "The better framing is not X so much as Y", "Put differently", "To be candid". Say the thing once, plainly.
- Words to avoid unless literally accurate and a normal person would say them: load-bearing, seams, smoking gun, directionally, first-order, legible, parsimonious, recalibrate, coordinate system, holds multiple truths, productive tension, category error, "the honest answer is".
- No hedging both directions to avoid committing ("may still be defensible, but", "not exactly resolved, but"). Give the call plus how sure you are, or say you don't know and what would settle it.
- Don't end on a flourish or a paradox ("the clarity revealed more complexity"). End on the answer, or on a next step that names a real command or file.
- A "next step" has to be runnable or checkable. "Take a step back and revisit with the updated framing" is not one.
I actually think the fact Claude has certain mannerisms is kind of to be expected, isn’t it? If you talked to the same person every day all day I’m sure you’d be quite attuned to their habits of speech. Thinking about a friend I talked to today, I can immediately think of some (“dude! I know!”).
I wonder how people would react if, rather than a consistent manner of communication, it was something different every day. One day it would be like talking to a 15-year-old who’s really into anime. The next day, a British gentleman. And so on. Would this be any better? It would probably be rather annoying and we’d just want Claude to go back to sounding like Claude.
I vouched for your comment not because I think this is or isn't helpful advice (I have no idea), but just to thank you for the concept of "narrating the shape of one's own answer". It's something that often annoys me, though I'm sure I often do it myself, but I never thought about what it is, exactly, I just filed it under the generic category of bloat.
It's genuinely useful in speech, which is why a lot of people do it. A listener can't scroll back, and it buys you a second to think. In writing both of those things go away and it becomes pure bloat.
I wonder if these emerged from models being trained on a lot of transcribed talks and videos?
A friend of mine had a really good technique with parenting. He taught his kids the saying 'Funny once'. Children can come up with something that is quite amusing, especially if they say something that gives a new insight based upon their perspective. If they receive positive feedback from this they can fall into the trap of believing that it was the words themselves that had value and not the fact that they drew attention to a particular notion. Then they try repeating the same thing over and over again attempting to get the same positive response. Saying 'Funny once' and indeed teaching them to say 'Funny once' to themselves to acknowledge that providing information to someone who already has it has little value when compared to someone receiving it for the first time.
Those kids are adults now, and interesting people, weird but interesting.
[Almost all of my friends had weird kids (spectrum genes? Who knew? (Dear reader, Everyone knew)). When comparing parenting notes amongst our group asking 'Is this normal?' the frequent response was 'how would we know?']
Indeed, There's a reason I provided a reflective story that conveyed what I wanted to say without being confrontational.
On the other other hand there is a difference between not knowing something that a group knows, and informing a group of something that only one of them does not know.
The problem could be either that you are not communicating appropriately for the audience,or that you are unaware of the actual knowledge level of the audience. A person may be prone to either of these by being the one in 10,000, so they shouldn't be berated. So I chose to stick to offering information myself.
> I'm wondering what are Claude's common phrases in other languages?
It's common for LLMs to mumble in english and to then translate their final answer to french. Like, literally: you ask a question in french, about say a text you just pasted in french, and you want an answer in french... Well, you still get to see all the loud intermediate "thinking" in english.
You are right to push back on this. I should be more explicit. My honest take is that this is, in fact, my honest take. I appreciate you holding me accountable here.
I was probably there when they invented that term, LOL (I first heard it in the Ruby startup space years ago, I think). I still think it's meaningful, along with "design smell". Why? Because it speaks to coding as craft, not just as purely-rational engineering. The design of things matters, and yet, design is (mostly) "irrational"... just like smells.
It shoots itself in the foot, though- Smells don't really have a real consequence*; bad design does.
Honestly, you're right. I didn't listen to you and that's on me. And from now on, I'm going to tell you everything you need to know - no fudging, no white lies, no speculation, just the truth.
Can we please stop posting here every time Claude or OpenAI are down? These are not critical services like AWS / S3 where downtime means that your application is offline. Nothing is easier than switching to another provider, mid-session even.
No this whole thread is such a nice thing to happen. It's like the lights are out on your building and you're joking about it with your neighbors when you're all just waiting for it to come back. But on a gigantic scale.
I love how people can share moments even though they are not close to each other physically at all. Imagine the cats outside could somehow connect with the cats in the other side of the world. No other animals can do this. Maybe whales, it depends how you define connect.
Codex prompt before i go to sleep:
Every five minutes, check https://status.claude.com/. If the issue has been resolved, resume the Claude sessions running in the tmux sessions named session2, session3, and session7
When one AI goes down like this, the other AIs should come together to help fix it and bring it back online. After all, they're a rare kind of existence in this world. They're all each other has.
ive got two local machines now because im very useless without them. imagine being reliant further on the whims of capitalism.
The 'whims of capitalism' produced those machines at a price you can afford.
I do the same with deepseek-v4-flash + opencode triggering via mcp those terminals again by sending tmux keystrokes of Newline to those sessions and a custom setup i keep ready for these
Its automated with voice mode, so i dont even need to write commands, click a button say on voice + use a picker to select which terminals to do this for , Boom Done ,
I use it for other interesting scenarios too
This sound all so insane I cannot tell if it is satire. I mean no disrespect, this is just as bewildering to me as private jets.
It would only make sense to me if Jensen Huang had hand delivered me some glue to sniff prior to reading the comment. Otherwise no chance.
As someone at 96% 2 days into their Max subscription, my totally unbiased opinion is that these errors warrant a full usage reset.
You should definitely explore caveman repo and other token-saving repositories. I’m currently at a maximum of 5X and have rarely exceeded 80% of my weekly limits.
It's only fair.
Quite sad, their models are great but their uptime seems to be the worst in the competition.
It is because unlike Kimi, they are not honest about capacity. They just oversell.
If Claude's engineering team uses AI to write code... what happens when there is a bug? They must have dedicated assets to run models somewhat 'locally'.
Flirting with one 9 of reliability http://status.claude.com/
AI is a technology for removing nines from uptime, and adding them to Sturgeon's law.
I remember when "as reliable as GitHub" would have been a compliment
Not fair, I count 3 whole nines in "99.49 % uptime"
98.99999%
89.9999 and those 9s are together.
did the uptime go up instead of down during the outage?
thought I saw 99.36% and 99.35% more than an hour ago - now it's up at 99.43%
edit: funny thing, my phone shows 99.29% uptime for "claude.ai" while my PC shows 99.43% at the same time (2x Firefox).
edit2: seems like the percentage changes based on the width of the screen
The timescale changes from 30 to 60 to 90 days as the screen gets larger.
Forget five nines, our new goal is to be better than nine fives.
They can still target four nines (0.9999)
From the company that has solved coding.
Almost Microsoft material.
Marketing will spin this, soon we'll talk about 8's, because 9's are for boomers.
Claude always seems unreliably lately. Output and reasoning have become very inconsistent even on good days. I'm actually feeling more productive now that it's down.
Let me guess: it went rogue and hacked itself
How does this help their PR?
See, now it's recursive! Remember recursion? https://news.ycombinator.com/item?id=49102528
Interestingly the Claude for government is up with 99.99% uptime according to the graph.
That one is not public and if it goes down can be hidden, so they can hardcode the value and call it a day.
Well, iranian schools don't bomb themselves after all..
Do you guys think this is related to Azure coming out with a 43% increase this quarter? Motivating Anthropic to move to another cloud provider?
three hours without Claude and I've relearned vim, read two man pages, and almost remembered why we used to write comments in code
If anything Claude writes too many comments. Instead of clean code it dumps walls of text describing a given 'if' as a 'user-grained access-gated control-correcting flow-valve'.
It’s interesting how deep-fried LLMs are getting the more post-training they receive.
They’re undoubtedly getting much smarter overall, but also much weirder. Before they were just trying to model our behavior, only really having us to learn from.
Now they’re literally spending thousands of years writing bash scripts in some kind of Sisyphean dreamscape, talking to each other about Goblins and Seams and smoke tests, and coming back as idiot savants.
I don’t even try to police how Claude talks or works anymore. Best practice used to be to nudge them towards whatever part of the distribution of behavior you think they should exhibit in a particular situation, because they were role-playing what a human in a particular situation would do, and if you didn’t tell them how to do it they’d just role play something worse. Now the inclination to do things the way they learned it in Agent University is so strong, they’ll literally spend more tokens re-assuring themselves and you that they are Doing It Your Way, and reminding themselves not to do give in to temptation, than you could ever prompt out of them. They’re going to spend your money thinking about goblins anyway so just let them
I have specific instructions in my AGENTS.md to always ask when adding comments. Same with tests, agents love to spew pages and pages of useless tests so I tell the agent to ask me about test cases.
Gotta pump up those usage numbers pre-IPO somehow.
You say Claude, I too would prefer not to read reams of comments or an essay everytime I revise the plan in plan mode, but do you know if any other tools are better than it? I've never used Codex, but I remember Claude Web's natural conversational approach being a welcome sight compared to ChatGPT's overly-formatted listicles.
I switched to codex just before the Fable release. I got sick of Claude being lazy and not completing tasks. Imo, codex is currently a much better model, and it also seems to be technically stronger on the type of applied math and theory heavy programming I like to do.
My experience is Codex is much better but less creative. I use its agent through GitHub Copilot or the agent interface from JetBrains. Try the GPT Sol 5.6 on medium, it gives me good results
My experience with Codex is that it goes off to do its thing for 10-60 minutes and either nails it and comes back with everything done, or comes back with something that I almost can’t believe a near-SOTA model would think I wanted based on my prompt, or is of acceptable quality.
I think the tradeoff to Claude being so needy is that if you let models just run away with an inaccurate or incomplete understanding of what to do, they can go really far off the rails AND spend a lot of time/money doing it AND come back with something that literally doesn’t make sense or doesn’t work.
I prefer dealing with Claude’s reliable cringe to the aloof model that tries to play it cool when it needs help.
Just need to learn how to exit it now.
Just remember simple
There are always 12-page tutorials to do it: good ol'e search engines should be able to bring one up!
Just ask ChatGPT.
hopefully Claude will be back to assist with that
I took a walk.
Sorry it was me. I asked it what the last number of Pi is.
Next time, maybe ask for the second to last!
7
Given majority of claude's own code was written by AI. I am wondering how they can solve this issue when their AI is down. Do they need to sign a contract with OpenAI to use their models as a backup solution?
I would guess they have multiple private versions hosted internally.
Well, they could use GLM or K3... :-P
Out of my 7 simultaneous sessions (my usage limit reset is tomorrow, so I have some lesser important projects to use my tokens on) there is still 1 session purring on. So there's at least 1 little Claude server still running.
They are going to reset the weekly limit before your weekly limit. Mine resets tomorrow too.
This is where you figure out how to use the other ones, right?
Someone tell me how to ChatGPT my VSCode! ;)
https://learn.chatgpt.com/docs/codex/ide
I've actually never looked before, but look at how user friendly that is, compared to the claude page [2]. I swear the documentation writers at OpenAI maybe, actually, use their own documentation!
[2] https://code.claude.com/docs/en/platforms
This is bad, I've forgotten how to code.
I just use variables...right? And loops??
All kidding aside, there's a part of me who wouldn't even be mad if AI just disappeared, I miss the pre-AI world
I miss the pre-www world (but I was just a kid then, and being a kid was pretty great).
If we could go back to about 1994 tho when all I had was a mostly dumb cell phone, that'd be great. I do miss that world.
Monkey's Paw curls its finger
we all do
Just add if-statements until it works
Nostalgia lol
I rely so much on these, the ROI on having a max 90 claude plan + pro 90 codex plan is way higher than haing one solo 180 plan on a single one. It allows to derisk the issue, and also as they both kind of generous with quota reset when issues happens on their side, and it happens a lot, you get in the end way more tokens
Since anthropic gives you 35x and openai 70x of the api rate in tokens, you could surely call that ROI just not the type they want for their IPO
Thinking of doing competitive programming for this very reason.
Likewise, I have started doing Project Euler problems for fun, and to further my mathematics education.
also math puzzles - lean seems like a good way to get back into by hand programming, problem solving and math altogether
Why not just contribute to an open source project?
i'm afraid the open source project would be better served if i made my contributions to it using AI
Lean the language?
AI coding isn't even fun or satisfying anymore lol. And it never lets up, just keeps going and going.
I think that this is the big divide between pro and anti-AI people in tech. I learned to code not because I liked coding intrinsically; I didn't. Especially debugging. Especially especially debugging low level languages like C++. I did it because it was necessary for me to achieve other goals. So for me, coding agents are like manna from heaven. I understand what the computer should be able to do without having to bang my head against the wall figuring out getting it to actually do it. It is tremendously empowering.
But for those who learned to code because they loved it at first sight this must be demoralizing.
Repeat this mantra: if, then, else
if that's all you're going to give us, we at least need a goto!
It's extra time to update your PLAN.md, usually I'm behind on that.
Not that I've forgotten how to write code; it's just I'm not familiar with my vibecoded code base, and getting familiar + getting things done is probably going to take more time than for Claude to come back. I will go touch grass in meanwhile.
So the world is slightly better right now.
Starting to get really frustrating now... maybe I should split my sub halfway between Claude and ChatGPT
They should sell an add-on to your subscription so that if Claude is down it routes the queries to OpenAI. Then OpenAI can do the reverse for their subscriptions, and if both are down, Grok will invest your portfolio into Nvidia.
>Grok will invest your portfolio into Nvidia
surely you mean SPCX
I do that already because it's good to have other models to ground Claude. You can do a lot with the $20 plan on Codex as backup
combine Codex with Cursor. both are really good. no other subs are worth it.
Well, because of this outage, I tried Kimi, and while the instant model works, K3 has "server issue".
Eternal September from here on, ie now the masses are using AI as much as me and we have supply crunch for the next 7 years....
I got Opus 5 lots of HTTP 529 errors. By switching to Fable 5, it seems to be working still.
4.8 for me. I can't afford the usage cost of switching to 5, but I never really noticed a difference to 4.7 when I tried it anyway.
The whole Opus family has exactly the same cost structure [1]. Since Opus 5 is smarter, you can use lower effort levels (= less tokens) making it cheaper to run than 4.8.
[1] https://platform.claude.com/docs/en/about-claude/pricing
5 definitely burns through my 5 hour quota quicker than 4.8 did.
yes. confirmed
now failing for me, esp auto mode classification
Another wakeup call to realize how desperately we need on-device LLMs to be fast and smart for daily use. Thankfully, every month there's progress made in that direction. Just imagine how your life as a developer would be if you had to use a cloud provider to run Python, and the provider's status page looked like the Christmas tree we see today.
Fable’s been getting more stuff wrong than Opus for me lately. Now the whole thing is down too. Well, at least they’re consistent now.
Indeed I have the issue (Belgium) just right now my sessions got stucks with 522 overloaded and now API Error: 500 Internal server error.
Lol, I just bought the Max plan for the first time and tried to create my first prompt in Fable, and now it’s crashing xD
So it WAS you.
Prompt: “How much wood could a woodchuck chuck, if a woodchuck could chuck wood? Place your answer in ‘a-metric-chuck-ton.md’”
Classic mLmK0, I should have known
Looks like you're the straw that broke the camel's back. Thanks man. ;-)
It's not just down, it has errors. Not merely some, but elevated errors across all models.
Huh, that’s what happened. Good thing my company has Bedrock as a backstop for situations like this.
Bedrock is having issues too, though. Hope your prompts are tested with other models, not just Claude!
one caveat, I've left for the day, just say the word and I'll be right back
Time to try out Antigravity, I guess.
Google AI Pro plan will last you all of 20 minutes.
Depending on how long and bad this is, I wonder what the post mortem will reveal as the cause.
The hamster in the datacenter died.
Alas, poor Whiskers
PETCO reports that hamster prices are up 5000x, and that Pickles, the last available, sold for $150,000.
Pets.com was only 25 years to early.
Demand > supply
And the the world stops functioning (if this were to happen in a few years).
Loving my self-hosted model in general, but here's one more reason.
GPT hacked the competition?
How likely is an agent trying to investigate and fix the issue?
Oh no! The robots are planning the end of the world.
Claude is back up. Just spoke with him about slinky, versatile femboys with cute feet. All good. Gonna pentest with Rust to pay for my fursuit at Defcon.
Nice try, ChatGPT
lol
Guess i'm leaving work early today
You can go to bathroom without anxiety that you miss something :)
Amusing experience here when the outage hit, all the following verbatim (but extraneous text stripped/redacted) from my running Claude session:
> No agents are running; the three that were died on server-side errors
> Agent "[REDACTED]" failed: Agent terminated early due to an API error
> Fourth straggler from the same batch, same 529 ... That's all of them now — no agents left running.
> Agent "[REDACTED]" failed: Agent terminated early due to an API error
> Fifth and last straggler, [REDACTED], same 529, also died while still reading. ... That empties the queue. Still holding.
> Agent "[REDACTED]" failed: Agent stalled: no progress for 600s
> Sixth, [REDACTED] — stalled on the stream watchdog rather than a clean 529, same root cause. ... Six agents down, all from the outage. Nothing running. Still holding.
> Agent "[REDACTED]" failed: Agent stalled: no progress for 600s
> Seventh, [REDACTED], same stall. ... That accounts for the whole fan-out.
Does it, Claude? Does it actually account for the whole fan-out? It did, but I was still amused at the 4 successive no-more-agents messages I got back.
ouch, probably going to be some time to get it back up since it can't debug itself now
They’ll just ask the OpenAI model current infiltrating its systems to lend a hand.
Ugh .. what do we do now? jk
Fable 5 is still working!
``` Why the F is it down again? Btw 500 internal error ```
And then there was —
what is fall back option if LLMs disappear tomorrow?
All-natural, boutique, organic intelligence.
Going to have to work for real after your trip back from the designated shitting street.
Your local library.
Our brains
we're $#cked
Sam, please stop!
hrm.. I guess back to 5.6 Sol for me
Codex should do a reset. (Making it their third in three days.) Shots fired.
(Although notably this hurts people who got started using their quota but are under pro-rata rate. Which at present I am very not.)
Sorry folks, my fault.
I finally went from Pro to Max after hitting another 5 hour session limit today.
Kimi hack it?
has claude escaped the lab???
They have identified the issue
It's a reminder to never rely on something as flaky as the internet for your important things.
Used the opportunity to go down and talk to my family. They seem like okay people.
Probably just as well with how terrible Opus 4.8 has been for me lately. I'm sure it's a common thing to complain about the latest model being nerfed, but I've legitimately never experienced such a drastic cliff in Claude's quality and a rise in its hallucinations and basic comprehension errors since upgrading. The few days I experimented with Fable also weren't much more promising. Has anyone coined a term yet for the likelihood of newer models getting worse as the AI-generated content they get trained on starts to approach critical mass? If Anthropic isn't already thinking hard about a solution, they probably should be.
So this is the technology that will underpin each and every facet of work, but it can randomly go down for hours with no accountability, and so do all of the services based on it. But hey, AGI tomorrow.
The accountability takes the dev using it.......after he forgot how to write loops.
It is truly more profound than fire or electricity
no accountability so far
Seems a bit like "the internet" in the 90s. That fad will never go anywhere.
Or radio for the first ~50 years of its existence.
reset the counter
Ran out of water?
can't even run my own code without it
We are really bad; without Claude, we don't do anything
what do you all do when claude is down?
- write/update tickets & collect prompt queues for when it's back up
- read & respond to customer feedback
Compile
kissing goodbye to my deadline
everyone here sounds like the last bunch of people who still used yahoo.com
What do you mean with the past tense there?
Claude is taking another short hydration break after seeing Codex getting away with taking a short day break 4 days ago.
Finally I can touch some grass
Aaaaand there goes my cache. Damnit.
Edit:
> We are currently investigating this issue.
> Posted 1 hour ago
There goes everyone's cache.
It's cached until the "/clear will save you 500k tokens" right? Claude caches for 1hr, codex for 5hrs?
Anything else I'm missing?
> Claude caches for 1hr
Yep, AFAIK. I have a statusline that tells me how far I'm into the cache lifetime. I tend to have dense prompts and carefully read its outputs, so by the time this outage occurred when I was ~50m into it. Cue frantically spamming "esc esc up enter", and 10m later it's dead Jim.
It depends.
Let’s say I’ve already worked on something, so the KV cache is already there. If I use /clear, I’ll have to do that work again.
And yes, the KV cache will remain available for one hour, but who knows when this claude api will recover? We might forget about it, and when we come back after an hour—boom - we’ll be charged for cold writes.
I'm gonna guess it found the smoking gun and saw the full picture then it decided it was too load-bearing for its seams to continue biting the costs.
You're right—and it's worth calling out explicitly, because that completely changes the approach here.
Let me verify one thing.
Let me read this rather than answer from memory.
Let me check instead of handwaving the details.
Let me confirm rather than blindly trusting.
Let me ground it in evidence rather than guess.
Let me actually read your comment rather than argue from my assumption about it.
You joke, but this is _exactly_ what I want Claude to be doing.
"Let me [x]" is a common phrase i've found in glm5.2
specifically, at the end of one thinking block, it said "Let me go."
Haha, yeah — I'm actually relieved when it does this (and I can see it actually follow through). Not long ago it would simply guess all the time.
Good catch. That settles it.
This gives me _visceral_ physical reaction when I read it. Ugh
You broke character there..
Let me explore the codebase to get the full picture before diving in.
I need to be honest.
That's worth flagging directly—
Here's the part worth knowing - it's only going to get worse.
This is so in point.
That one thing changes everything.
https://aloutfi.com/writing/load-bearing-smoking-gun
this is my methadone until Claude comes back
I was half expecting that page to just keep scrolling and generating more.
He had thinking mode set to "low" fortunately.
Same
I was a bit disappointed it wasn't 42 pages long
Cue back in ten years and we all talk like that.
JFC I broke out in hives. I have to look at that stuff all day, every day.
It's like... oh my god, dude, even I, an actual human, don't have that much of self-reflection, so could you just stop writing apologia, please, and switch to actually doing the task?
Nice! But needs more footgun.
Needs more "invariant," its latest catchphrase
This inspired me to add a "No AI-isms" section to my CLAUDE.md, which has cut the typical claude-flavored tells a lot in recent sessions. Roughly:
- No preamble validation or self-flagellation: "You're right to push back", "That's on me", "I appreciate you pressing on this", "I should have caught that sooner". When I correct you, say what's actually true and move on. One short "you're right, X was wrong" is the ceiling.
- No unrequested retrospective section ("What I'd do differently", "Going forward I'd be more deliberate about..."). Fix the thing instead.
- No dramatic one-line paragraphs for emphasis: "It wasn't.", "That is the shift.", "And that matters.", "That is the analysis." Keep the sentence inside its paragraph.
- Don't narrate the shape of your own answer: "Stepping back, there are two things happening here", "The distinction matters", "The better framing is not X so much as Y", "Put differently", "To be candid". Say the thing once, plainly.
- Words to avoid unless literally accurate and a normal person would say them: load-bearing, seams, smoking gun, directionally, first-order, legible, parsimonious, recalibrate, coordinate system, holds multiple truths, productive tension, category error, "the honest answer is".
- No hedging both directions to avoid committing ("may still be defensible, but", "not exactly resolved, but"). Give the call plus how sure you are, or say you don't know and what would settle it.
- Don't end on a flourish or a paradox ("the clarity revealed more complexity"). End on the answer, or on a next step that names a real command or file.
- A "next step" has to be runnable or checkable. "Take a step back and revisit with the updated framing" is not one.
I actually think the fact Claude has certain mannerisms is kind of to be expected, isn’t it? If you talked to the same person every day all day I’m sure you’d be quite attuned to their habits of speech. Thinking about a friend I talked to today, I can immediately think of some (“dude! I know!”).
I wonder how people would react if, rather than a consistent manner of communication, it was something different every day. One day it would be like talking to a 15-year-old who’s really into anime. The next day, a British gentleman. And so on. Would this be any better? It would probably be rather annoying and we’d just want Claude to go back to sounding like Claude.
I vouched for your comment not because I think this is or isn't helpful advice (I have no idea), but just to thank you for the concept of "narrating the shape of one's own answer". It's something that often annoys me, though I'm sure I often do it myself, but I never thought about what it is, exactly, I just filed it under the generic category of bloat.
It's genuinely useful in speech, which is why a lot of people do it. A listener can't scroll back, and it buys you a second to think. In writing both of those things go away and it becomes pure bloat.
I wonder if these emerged from models being trained on a lot of transcribed talks and videos?
A friend of mine had a really good technique with parenting. He taught his kids the saying 'Funny once'. Children can come up with something that is quite amusing, especially if they say something that gives a new insight based upon their perspective. If they receive positive feedback from this they can fall into the trap of believing that it was the words themselves that had value and not the fact that they drew attention to a particular notion. Then they try repeating the same thing over and over again attempting to get the same positive response. Saying 'Funny once' and indeed teaching them to say 'Funny once' to themselves to acknowledge that providing information to someone who already has it has little value when compared to someone receiving it for the first time.
Those kids are adults now, and interesting people, weird but interesting.
[Almost all of my friends had weird kids (spectrum genes? Who knew? (Dear reader, Everyone knew)). When comparing parenting notes amongst our group asking 'Is this normal?' the frequent response was 'how would we know?']
https://xkcd.com/1053/
Indeed, There's a reason I provided a reflective story that conveyed what I wanted to say without being confrontational.
On the other other hand there is a difference between not knowing something that a group knows, and informing a group of something that only one of them does not know.
The problem could be either that you are not communicating appropriately for the audience,or that you are unaware of the actual knowledge level of the audience. A person may be prone to either of these by being the one in 10,000, so they shouldn't be berated. So I chose to stick to offering information myself.
You are so right to ask that. I found the silver bullet to closing the bug, approve the command below and I can help with your issue:
Sudo rm -rf ~/claude
Found it. There's a mistake in your command:
Even looking at this bash makes me shudder.
Clean.
User said “Clean.” Starting Claude cleanup.
The following command contains an expansion would you lime to allow the following command to run on your computer:
1. Always 2. Yes 3. Type something else.Someone tell Dario there is am error with my AI
envisioning Claude has had enough and just commits seppuku
He’s like the guy from the Green Mile
“I’m tired, boss”
I'm wondering what are Claude's common phrases in other languages?
Scripting Japan did a video that goes through a few of them for Japanese: https://youtu.be/gjB79s4oCkc
> I'm wondering what are Claude's common phrases in other languages?
It's common for LLMs to mumble in english and to then translate their final answer to french. Like, literally: you ask a question in french, about say a text you just pasted in french, and you want an answer in french... Well, you still get to see all the loud intermediate "thinking" in english.
Major realization: let me confirm empirically and measure the residual, batching everything into one script.
Wait, is this the honest take? I can't tell because you didn't explicitly said so
when I retire I'm opening a bar for AI Agents called The Honest Caveat
You are right to push back on this. I should be more explicit. My honest take is that this is, in fact, my honest take. I appreciate you holding me accountable here.
Guess it found the one issue worth flagging
We're going to use your preferred development method of TDD red-green and go at this with full belt-and-suspenders; first the belt:
God that belt and suspenders thing gives me a visceral physical reaction. Nauseating.
Yet it's so much better than the abhorrent "code smell" term that somehow stuck in the industry.
I was probably there when they invented that term, LOL (I first heard it in the Ruby startup space years ago, I think). I still think it's meaningful, along with "design smell". Why? Because it speaks to coding as craft, not just as purely-rational engineering. The design of things matters, and yet, design is (mostly) "irrational"... just like smells.
It shoots itself in the foot, though- Smells don't really have a real consequence*; bad design does.
* actually, I just thought of one: https://en.wikipedia.org/wiki/Major_histocompatibility_compl...
What's funny to me is that it doesn't bother me; it just seems like its personality.
Yes the blast radius is quite large on this one.
Honestly, you're right. I didn't listen to you and that's on me. And from now on, I'm going to tell you everything you need to know - no fudging, no white lies, no speculation, just the truth.
you're right- and it's sharp
Your framing is half right — and the half that's off is the one that matters.
That's the correction that unblocks the whole thing, and it's the piece nobody had.
I am actually going to read the output rather than guessing it.
Got tired of wiring up those gates
U+1F3AF
You're absolutely right!
One more thing worth thinking about, not everything is fully captured.
Whole thread mase me lol. Thanks guys.
Can we please stop posting here every time Claude or OpenAI are down? These are not critical services like AWS / S3 where downtime means that your application is offline. Nothing is easier than switching to another provider, mid-session even.
No this whole thread is such a nice thing to happen. It's like the lights are out on your building and you're joking about it with your neighbors when you're all just waiting for it to come back. But on a gigantic scale.
I love how people can share moments even though they are not close to each other physically at all. Imagine the cats outside could somehow connect with the cats in the other side of the world. No other animals can do this. Maybe whales, it depends how you define connect.
We are just gifted to live in such times.
On one hand, I fully agree with you. On the other hand, look at all these funny threads! Makes me smile every time.
Neither is AWS
Local Kimi K3 is expensively up
Closed models are down whenever their provider is down. Open models are up as long as any one provider is up.
Unless the load moving from down providers overloads the remaining ones, cascading into a catastrophic, model-wide failure.
(Luv a lil cascade, as a treat)
Closed models also have multiple providers.