Over the last two days, I have observed a significant degradation in Cloud Sonnet 4. Today, I checked the model, and it replied that it is version 3.5. Model 4 thinking identifies itself as 4. Two days ago, the results were fantastic, but today, its output is terrible! I added a screenshot of how the model identifies itself.
I checked all the models, and they identify themselves correctly but only sonnet 4 shows it 3.5.
This has been debunked as false response by Claude so many times in the forum.
Anthropic does not give the models an number, they are just called âClaudeâ.
Since Claude 3.5 Sonnet was in the training information of Claude 4 it is known and sometimes Claude 4 hallucinates the name.
Danke schön
Then why sometimes it shows me âclaude-3.5-sonnetâ even when I select âo3â? Do you say that itâs also because claude models were in training data for gpt-o3?
Could you please post a request ID that can be analyzed by Cursor Team (privacy disabled for that request where you reproduce it)? I have never heard it happening cross model before so only a proper analysis can address your question.
Cursor does not route o3 requests to Claude 3.5 sonnet.
This still doesnât change the fact that Claude 4 acts more like chatgpt with emojis and only providing broken text (unless my setup is just wrong)? I have only got claude 3.5 to work (Only referencing anthro LLM)
Thanks for your reply. When you make statements like âonly providing broken textâ without a proper example or a full bug report there is nothing I can review to give you feedback or ask Cursor Team to investigate a bug.
Therefore I sincerely suggest you to file a full bug report. Issues like you describe should be analyzed and fixed where possible. Iâm also offering you to tag me with @ into your bug report so I dont miss it among hundreds of posts I handle every day.
To clarify, it would not pertain or stay in boundaries within my explicit rules given. I had much more success with Claude 3.5 with more intuitive task or comprehensive ones than I would with Claude 4. To expand on this, Claude 4 would only provide emojis even when told not to not to mention the lack of âthoughtâ or âlogicâ was put into the coding even when the task was at bare minimum relatively âeasyâ. Claude 4 produced files that were only TODO or placeholders etc. It was more of a general experience than a specific one so its harder to pinpoint exactly what is the root cause, however I just know that a ânewerâ model should be âbetterâ?
Right, I had the same issue when switching from Sonnet 3.5 to 4.
My prompts, rules and documentation that worked so well for 3.5 didnt work at all with 4 (or 3.7). I tried different approaches and found that 4 does work well with many details and so strict rules or requirements. As it is a hybrid reasoning model it needs a bit different prompting.
So I removed all rules, all detailed requirements, and wrote following as rule:
Follow SOLID, DRY & SRP.
(Mentioning the programming language and framework can help as well)
Next I wrote a basic features.md file in docs/ for new feature to implement. Literally just the requirements without saying what where or how needs to be changed, and asked 4:
Analyze requirements in docs/feature_name/feature MD, check existing code to see what needs to be changed and create an implementation plan as MD file in same folder.
After that it created a quite good implementation plan, I gave it some adjustments, with short reasons why this is important.
Adjustments:
- Change ⊠to ⊠as this improves stability
Next I prompt ask it to implement the changes
Implement features in implementation plan in docs/feature_name and update there a progress.md file.
After that sometimes adjustments are required, but mostly over 90% is really correct.
Sometimes I just feel like they have a mind of there own XD
