claude-code - 💡(How to fix) Fix Opus 4.7 regression vs 4.6: ~2x token burn, mid-session stalls, no quality gain

TL;DR

Reverting to Opus 4.6 may mitigate the regressions observed in Opus 4.7, including no perceptible capability gain, increased token consumption, mid-session stalls, and UI/harness regression.

Guidance

• Investigate the differences in model behavior between Opus 4.6 and 4.7 to identify the root cause of the regressions.
• Compare the system prompts and harness changes introduced in the UI refresh to determine if they are contributing to the model's worse behavior.
• Test Opus 4.6 with the same tasks and inputs as Opus 4.7 to verify that the regressions are specific to the newer version.
• Consider collecting more data on the mid-session stalls to understand the patterns and triggers of this issue.

Example

No code snippet is provided as the issue does not involve specific code changes.

Notes

The investigation should focus on the changes introduced in Opus 4.7 and the UI refresh, as these are the most likely causes of the regressions. The fact that Opus 4.6 is still available as a fallback suggests that it may be possible to revert to the previous version while the issues with Opus 4.7 are being addressed.

Recommendation

Apply workaround: Keep Opus 4.6 selectable as a stable fallback while Opus 4.7 is being addressed, as this will allow users to continue using a version that is known to work without the regressions.