Hey team, I’m using Claude 4.5 Sonnet and all of a sudden it’s giving me this error message. It has been working flawlessly previously for the last couple of weeks (it worked fine earlier today as well). After getting the error message I tried to increase the output length and it’s still not working. Any thoughts on this?
Is anyone else using Claude 4.5 Sonnet getting this error?
I tried to switch from my API over to the pickaxe credit system and the same error came up. I don’t see anywhere to access the “thinking budget tokens” through their UI. Thanks guys!
That’s a nice report. It looks like the immediate solution is to increase your max output tokens setting. @nathaniel please don’t build SkyNet, because we know that doesn’t end well, but maybe a tiny bit of self-awareness by the platform? I see a related thing when I am generating 3 images…2000 bytes isn’t enough. I think the new Builder usually surfaces such issues eventually, but it would be pretty magical to get a little toast : ‘Hey, we noticed you are outputting a lot of data…seems like we should increase the Max Output Tokens setting to around #####”
The issue is that your max output (2050 tokens) is too low relative to the thinking budget being used.
Solutions:
Increase max output length to at least 3000-4000 tokens minimum
Or disable extended thinking if not needed
Or explicitly set a lower thinking budget (minimum 1024 tokens, must be < max_tokens)
The thinking budget must always be less than max_tokens, with a minimum of 1024 tokens.
Quick fix: Set max output to 4096 tokens.
Also, in your screenshot, the memory buffer is set too low. Set it to at least 5x your max output length (minimum 10,000 tokens for a 2050 max output).
What memory buffer does: Retains the last N tokens from conversation history.
Current problem:
Each exchange uses ~3,550 tokens (1,500 input + 2,050 output)
Memory buffer = 1,025 tokens
The AI can only recall the last 1,025 tokens (~29% of one exchange)
It will forget context after just one turn
Result: Severe amnesia. The chatbot cannot maintain context across multiple messages.
Never mind I spoke too soon. It’s throwing the error again.
I did raise the token limits, and the error went away for about an hour, but then it came back — and now I’m getting the same message on every Claude model (4.5 Sonnet, 4.1 Opus, 4 Opus, 4 Sonnet, 3.7 Sonnet):
Provider Error:max_tokens must be greater than thinking.budget_tokens.
Here’s a concise recap of what I’ve tried:
Token settings
Max Output raised stepwise: 2050 → 4096 → 8192 (also briefly 32000 max to test it and it still didnt work).
Max Input now 1500.
Memory Buffer now 1500 (previously higher; I even tried this at 32000).
Model + session isolation
Switched to GPT-5-Chat → works perfectly.
Switched back to Claude 4.5 Sonnet or 4.1 Opus → error returns.
Tried all Claude variants listed above → same error on all.
Published after each change and tested in brand-new chats (including incognito).
It does however work in the pickaxe preview window (this is the only place it works, as soon as I publish and go to a browser or phone it throws the error.
I tried turning off my API and using pickaxe credits and the same problem occurred.
Also disabled all actions (Google Search + Nano Banana) and still saw the error.
Prompt & memory hygiene
Added a runtime rule: “do not use extended/hidden reasoning.”
Tightened output length in the prompt and trimmed user memory for heavy threads.
Given the error wording, it seems like the Anthropic route is still getting a thinking block (with budget_tokens) attached somewhere in the chain, even when using models that shouldn’t auto-enable it and even with actions off. I dont see anywhere in the pickaxe UI to change the value of the budget_tokens or to turn off reasoning.
Hey @motheraigaia I referred to Anthropic documentation when answering you first post so that’s the official way to deal with this specific issue you’re facing.
Since the issue still persists on your end then it’s likely an issue from how pickaxe configured their API key.
Btw, if you haven’t already joined the Pickaxe Prospectors Academy on Skool - you’re welcome to apply! We’re all Pickaxe community members and experts learning from each other. I offer free Pickaxe training, weekly updates, and build tips/tricks. Regardless of what your goals are, you can learn how to DIY and get a LOT of resources in the process!
In the exact same boat! I could’ve written this post. I so appreciate the follow up conversation and I will give those things a try. From what I’m reading it sounds like the problem is with the Claude interface which I am also using.
First check is just to switch that over to a GPT learning model and see if the problem goes away.
I also have no problem when using in the pickaxe preview screen. It’s only when I go to my studio that I’m getting that exact error message.
Well switching from Claude 4.1 over to GPT-5 got rid of the error immediately. I’ll leave it like that until we get a clear message that it isn’t happening anymore with Claude (which I prefer for the output my community needs).
Thank you for reporting these errors and for sharing your experiences in detail. The Claude-related issues you are seeing have been reported to our engineering team. They are currently investigating the cause.
We appreciate your patience and for helping us identify this issue.
Hey Barry, I’m glad I’m not the only one experiencing the problem. I’ve been going crazy trying to get it to work. Currently it’s still not solved. I emailed the pickaxe team, hopefully they can get it figured out pretty soon because all of the Claude models are unusable at this point.
You’re welcome. I’ve set this issue to high priority, and we’re hoping to have a fix soon. Our engineering team mentioned that these errors are most likely the result of recent changes Claude made to their API without prior notice. We sincerely apologize for any inconvenience this may have caused.
In the meantime, I recommend switching your tools to a non-Claude model so they continue functioning properly. Thank you for your patience and understanding while we work on resolving this.
We’ve just pushed some updates, and this issue should now be resolved. Please try using Claude models within your tools again.
If you continue to experience any problems or notice unexpected behavior, don’t hesitate to reach out. We appreciate your patience and your valuable feedback as we work to enhance your experience.
Yeah unfortunately the error is still there for me too. I tried switching between some other models to maybe get it to reset but nothing has worked. We’re about 4 hours after the new changes and the problem still persists. Thanks to the team for quickly working on it. I have faith you guys will figure it out. Thanks @raspyni@danny_support
Thank you for letting us know that the issue is persisting. I’ve shared the video and your feedback with our team, and we are actively investigating.
In the meantime, please switch to a different model so you can continue using your tools without interruption. We appreciate your patience and understanding while we work on resolving this.
You bet and THANK YOU for being so on it. @motheraigaia I’m now using Grok4 Fast and LOVING the output it’s giving my users. Grateful that Pickaxe gives us access to so many to try - especially at times like this where the one I’ve been using glitches.
Hey Barry Sonnet 4.5 appears to be working again for me, I’ll keep you updated if it goes back down. Let me know if it’s working for you. Great work team @danny_support !
Thank you both for your continued feedback and for sharing your findings as we worked through this issue. Your input has been a valuable part of improving the platform. If the issue comes up again or you notice anything similar with the Claude models, please don’t hesitate to reach out.
!