That rate limit improvement people are noticing? It's not Anthropic raising limits. It's Anthropic kicking ~135K OpenClaw instances off subscription credits on April 4.
OpenClaw sessions bypassed Claude Code's prompt cache, so each one was burning 10-20x more compute per task. A single instance could rack up $1K-$5K/day in API-equivalent compute on a $200/mo Max plan. 60% of active sessions were running on subscription credits, so that's a massive amount of load just... gone from the shared pool.
Google did the same thing to AI Ultra subscribers the same week, but without any warning. Anthropic at least gave notice.
The real problem is that flat-rate subscriptions for AI compute were always a gym-membership bet. Sell to people who won't max it out. Autonomous agents broke that model completely. This won't be the last repricing.
I run Claude Sonnet 4.6 via Github Copilot and it seems very reasonable to me there.
I just create an issue and assign it to Copilot and then hop into its session and sometimes redirect or give feedback after it reaches a stopping point and I've had the chance to pull it down and test it. I'm closing out 2-3 semi-complicated features a day on it in my off work hours right now for my personal projects and I didn't even get close to hitting the cap for the $10/month I'm paying for it right now (although each month it is creeping up as I start doing more and more with it). And I'm still getting way more done than I was when I was coding it all manually before these models.
One of the things I'm making with it right now I can't even sell (or probably even make public), I just want to play my favorite deckbuilding card game (that has lots of different cards with different effects) on my mobile and there isn't a good version of it, so I'm trying to vibe code it into existence (and have gotten pretty far along on it, most of the core game rules and about a quarter of the card effects are implemented right now). I'm pretty close to able to play a full game of it with a limited set of cards already. The presentation is mostly text but it gets the job done.
Work uses Codex within Visual Studio Code and that I got close to hitting the monthly limit on, but I haven't yet.
I get GitHub Copilot Pro for free for some reason. One day I checked and it was just there. So I use that until it runs out. When it does, https://synthetic.new with Kimi K2.5 works surprisingly well for small tasks where I still make all the decisions.
But I find no matter what I use, it still makes more sense to code by hand for anything that actually matters.
The things I've vibe coded are throwaway scripts to generate a gif, user scripts to tweak annoying websites, and various utilities that just need to work.
I have the $100/mo Claude plan, I've used 5% of my weekly and it resets this evening. I'm not a heavy user, but I also feel like I'm not a slouch either. I don't get how people are rolling through their usage so fast.
I should probably drop to that plan. I'm averaging around $800/mo in token usage based on ccusage, but I never hit plan limits and am told to wait. I've used it quite extensively this week with a lot of changes to local infrastructure, but still showing 0% utilization across current and weekly sessions according to /usage.
Also worth checking if you're running long agentic loops — each tool call in a multi-step task counts against the window independently. So before switching providers, disable the extended context and run a day. It's probably not the model.
I moved to GLM-5.1 with their coding plan. It's better than both Opus and Sonnet, while costing a ton less. But it does slow down sometimes during peak hour, but I'm a patient man.
Last 2 weeks I was using it more or less all day on Opus running skills to write PRDs and then code and tests to solve the PRDs, never hit the session limit.
Last 2 days I hit the cap in about an hour of kicking off my skills workflow.
On the paid enterprise team plan this is really bad.
Well, maybe this is an unpopular opinion, but I prefer the Gemini Cli, I paid Google AI Pro for the year and it is perfect for me, even though it's true that pro model sometimes takes like 2 - 4 minutes to answer
That rate limit improvement people are noticing? It's not Anthropic raising limits. It's Anthropic kicking ~135K OpenClaw instances off subscription credits on April 4.
OpenClaw sessions bypassed Claude Code's prompt cache, so each one was burning 10-20x more compute per task. A single instance could rack up $1K-$5K/day in API-equivalent compute on a $200/mo Max plan. 60% of active sessions were running on subscription credits, so that's a massive amount of load just... gone from the shared pool.
Google did the same thing to AI Ultra subscribers the same week, but without any warning. Anthropic at least gave notice.
Full analysis with 16 citations here: https://sloppish.com/the-borrowed-cloud.html
The real problem is that flat-rate subscriptions for AI compute were always a gym-membership bet. Sell to people who won't max it out. Autonomous agents broke that model completely. This won't be the last repricing.
I run Claude Sonnet 4.6 via Github Copilot and it seems very reasonable to me there.
I just create an issue and assign it to Copilot and then hop into its session and sometimes redirect or give feedback after it reaches a stopping point and I've had the chance to pull it down and test it. I'm closing out 2-3 semi-complicated features a day on it in my off work hours right now for my personal projects and I didn't even get close to hitting the cap for the $10/month I'm paying for it right now (although each month it is creeping up as I start doing more and more with it). And I'm still getting way more done than I was when I was coding it all manually before these models.
One of the things I'm making with it right now I can't even sell (or probably even make public), I just want to play my favorite deckbuilding card game (that has lots of different cards with different effects) on my mobile and there isn't a good version of it, so I'm trying to vibe code it into existence (and have gotten pretty far along on it, most of the core game rules and about a quarter of the card effects are implemented right now). I'm pretty close to able to play a full game of it with a limited set of cards already. The presentation is mostly text but it gets the job done.
Work uses Codex within Visual Studio Code and that I got close to hitting the monthly limit on, but I haven't yet.
I get GitHub Copilot Pro for free for some reason. One day I checked and it was just there. So I use that until it runs out. When it does, https://synthetic.new with Kimi K2.5 works surprisingly well for small tasks where I still make all the decisions.
But I find no matter what I use, it still makes more sense to code by hand for anything that actually matters.
The things I've vibe coded are throwaway scripts to generate a gif, user scripts to tweak annoying websites, and various utilities that just need to work.
I have the $100/mo Claude plan, I've used 5% of my weekly and it resets this evening. I'm not a heavy user, but I also feel like I'm not a slouch either. I don't get how people are rolling through their usage so fast.
I should probably drop to that plan. I'm averaging around $800/mo in token usage based on ccusage, but I never hit plan limits and am told to wait. I've used it quite extensively this week with a lot of changes to local infrastructure, but still showing 0% utilization across current and weekly sessions according to /usage.
Check under settings on the website.
I can only assume they're eother setting it to Opus all the time, or they're using something like Ralph Wiggum.
I burned my week quota working on one small repo (with a lot of data files though) for one working day yesterday. It wasn’t like that before.
Something definitely changed, or it’s somehow reading all that data over and over again
Turn of the 1M context that got enabled by default. Long sessions eat through the tokens much faster.
Your sessions were probably getting auto-compacted much earlier before the context window got larger.
Also worth checking if you're running long agentic loops — each tool call in a multi-step task counts against the window independently. So before switching providers, disable the extended context and run a day. It's probably not the model.
Codex, it's much more generous. And doesn't lock you into using their CLI.
Still I'm a bit surprised you burn through tokens that quickly. I rarely ever reach my limit.
I moved to GLM-5.1 with their coding plan. It's better than both Opus and Sonnet, while costing a ton less. But it does slow down sometimes during peak hour, but I'm a patient man.
It feels lower.
Last 2 weeks I was using it more or less all day on Opus running skills to write PRDs and then code and tests to solve the PRDs, never hit the session limit.
Last 2 days I hit the cap in about an hour of kicking off my skills workflow.
On the paid enterprise team plan this is really bad.
Very good overview of options here: https://ai.davis7.sh
Explained in more detail: https://youtu.be/1WFgIjAvMDw?t=882
TL;DR:
- Cursor Ultra
- OpenAI Codex
- OpenCode Black (currently not accepting new subs)
> - OpenCode Black (currently not accepting new subs)
Temporarily paused on new subs. It'll be back.
Well, maybe this is an unpopular opinion, but I prefer the Gemini Cli, I paid Google AI Pro for the year and it is perfect for me, even though it's true that pro model sometimes takes like 2 - 4 minutes to answer
[flagged]
[dead]
[dead]
[dead]
[dead]
[dead]
[flagged]