[BUG] On Max 20x, Opus gets downgraded to Sonnet at 20% usage in 5 hour window.

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The "Get Help" AI in Claude for Windows desktop application insists that since I upgraded to Max 20x I should not get downgraded silently to Sonnet from Opus until I hit 50% usage limits. I was on Pro, upgraded to Max 5x, was getting downgraded at 20%, eventually decided to try Max 20x after the "Get Help" AI said I could get 4x tokens and not be downgraded until the 50% mark. But I've designed projects and features with Opus and hit 20% and told it to work on it and Sonnet took over and disregarded all the design, decided it would "simplify" things and combine things that were not related into the same MCP server, when I had designed it with Opus to be two separate MCP servers. I woke up hours later to find that we had to revert all the changes and have Opus start all over in another session. As soon as I hit 20%, the usage increased at a rate of 5 times slower, because we were using Sonnet instead of Opus. Sonnet cannot follow simple instructions. The only way I can use Sonnet to write code is if I have Opus write the prompt for it and review the results afterwards to verify it was done correctly or it has to revise the prompt and have Sonnet redo it or just do it itself. I'm on Max 20x and I can only use 20% of my usage in any 5 hour window. I'm wasting 80% of my time and tokens. And I've spend a lot of my time trying to teach Opus how to work around the 20% limits on Max 20x. And the "Get Help" insists I'm experiencing a bug.

What Should Happen?

According to "Get Help" AI, I should be able to hit 50% usage limits before I get downgraded to Sonnet. In my opinion, I shouldn't ever get downgraded to Sonnet and I should be allowed to use 100% of my tokens using Opus and having Opus delegate and fire off Agents when Sonnet can do some of the work.
It keeps giving me CLI commands I don't have access to and telling me to use those. I cannot always be around at 20% and I have to stop at 20% in any 5 hour window because of this. I've had to re-audit my code 3 times now, (which used a lot of Opus tokens) because I didn't realize I was being downgraded in the middle of the session while Opus was writing code for me. I lost several hours of sleep last night trying to fix what Sonnet broke and rework all the training I did for Opus to know how to work around the downgrades. It was fixing the code all day today. I had to wait 3 1/2 hours to midnight only to find I had to wait another 20 minutes in order to continue working on my projects. You should seriously upgrade your "Get Help" AI to Opus as well. I recommend re-allocating all your servers for Haiku and Sonnet to be used for Opus; drop Haiku and Sonnet models and retire them. They are useless for coding. Opus would do a much better job for all the other tasks people use Haiku and Sonnet for anyways. And since all the hardware you already own for Haiku and Sonnet could be reallocated for Opus, you could maybe let people use it more and charge less for it.

Error Messages/Logs

Steps to Reproduce

Wait for your 5 hour window to start... Use Opus until you hit the 20% mark... use ccusage (Node.js) to see your token usage... continue using Claude without changing the model. After the same amount of new token use, your %usage only goes up by 5 or 6% so instead of being at 40%, you are at 25-26%. And then wait until the next 5 hour window starts and have Opus review all the work that was done after you hit 20% and it will tell you it's not up to par. It can tell it was Sonnet, just like I can tell it was Sonnet.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Version 1.5354.0 (9a9e3d)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

I'm using Claude for Windows which I downloaded from claude.ai