Developers say Claude Opus 5.5 got worse, but a test shows no change
In logs from one developer, 25.4% of answers needed a correction after October 19, against 5.2% before.
Developers say Claude Opus 5.5, the coding model that Anthropic released on September 22, fails at simple edits. They say it repeats rejected solutions and ignores files that it read. The complaints peaked from November 4 to 6, and Anthropic gave no answer as of November 7. Opus 5.5 still leads most November rankings of top models.
What developers report
Developers say the model acts as Opus 5 does. One adjustment can change the result. Documents of Anthropic say Opus 5.5 uses medium effort by default, one level lower than Opus 5.
The numbers
One heavy user checked 6,638 answers in local Claude Code logs. Before October 19, 5.2% needed a correction. After that date, 25.4% needed it, with the same setup. LiveNerf operates the same test set each day and reports no change after the first week.
What is not known
No one proved a drop in quality, and no one ruled it out. The logs come from one user. Anthropic did not answer the complaints.
Insight: A top score does not prove the model you use
On October 12, Artificial Analysis gave Opus 5.5 the top score of 52% on FrontierCode 1.1. A November study gave the model more data. Its market gains in double auctions fell from 94% to 64%. A score fits one setup, and a vendor default can change the model that a buyer uses.
Sources
Posted
