Noticeable capability degradation in Claude Opus/Sonnet 4.6 after resolving #190758

[YKDZ]

🏷️ Discussion Type

Question

💬 Feature/Topic Area

VS Code

Body

Hi everyone,

Ever since the rate-limit issue reported in #190758 was resolved, I've observed a noticeable drop in the reasoning capabilities of both Claude Opus 4.6 and Claude Sonnet 4.6 (with Sonnet being particularly obvious) when using them inside VS Code over the past couple of days.

Specific issues I've noticed include:

• Decreased instruction-following capability.
• A significant decline in the quality of the generated code.

Has anyone else experienced this recently? Also, are there any known user-side workarounds for this situation?

As a side note, I am a Pro+ subscriber, and I am quite dissatisfied with this unannounced degradation in service quality.

[Synalix]

I haven't noticed this myself but you're not the only one reporting it. It could be a server-side change on Anthropic's end rather than anything you can fix locally.

A few things worth trying in the meantime, switching between Sonnet and Opus to see if one performs better for you, restarting the Copilot extension, or checking if there's a newer version of the VS Code extension available. Other than that it's probably worth waiting to see if GitHub/Anthropic acknowledge it.

[Priyanshu-byte-coder]

You are not alone in noticing variability in model performance over time, especially when providers adjust rate limits, routing logic, or model configurations. Changes made to improve stability or reduce latency can sometimes affect how consistently the model follows instructions or generates code.

A few possible reasons for the perceived drop in reasoning or code quality:

• Model routing changes – requests may be handled by slightly different variants or configurations optimized for speed or cost.
• Context window handling – truncation or compression of long prompts can affect instruction-following accuracy.
• Temperature or decoding adjustments – small parameter changes can noticeably impact determinism and reasoning style.
• Extension-side changes – VS Code extensions may update prompt formatting, system instructions, or tool usage behavior.
• Caching or fallback models – under heavy load, providers sometimes temporarily route to fallback configurations.

Things you can try on the user side:

1. Reduce prompt ambiguity – structure instructions into clear numbered steps.
2. Split complex requests into smaller sequential prompts.
3. Explicitly specify constraints (language version, libraries, expected output format).
4. Reset the conversation/session to avoid degraded context accumulation.
5. Check whether the VS Code extension has updated recently.
6. Compare results between “chat” and “edit” modes if available.
7. Try slightly more explicit system-style instructions (e.g., “Follow all requirements exactly”).
8. Monitor provider status pages or changelogs for silent configuration updates.

If multiple users report similar behavior in the same time window, it often indicates backend tuning rather than an account-specific issue.

Providing short reproducible examples (prompt + output differences) can also help maintainers or providers identify whether a regression has occurred.

Hopefully this is temporary while configurations stabilize after the rate-limit changes.