LiveNerf finds no measurable Opus 5.5 decline after 47 days
AI summary
The LiveNerf project reports that 47 consecutive days and 3,666 samples found no sustained, measurable decline in Opus 5.5 performance under its benchmark. Its original 10-day baseline averaged 58.4%, while the latest daily result was 61.5%, although results varied by day. The project also says identifiers seen in Claude Code may reflect routing or configuration differences, which do not by themselves demonstrate reduced underlying model capability.
In-depth insight
The results provide a concrete independent check against claims that Opus 5.5 was weakened, while distinguishing possible changes to the surrounding product stack from changes to the model itself.