Opus 5.5 degraded? Users suspect quality drop after launch

Shortly after Anthropic launched the AI model Claude Opus 5.5 on September 22, 2026, claims have surfaced that it is getting worse over time. So far these are only user impressions, and no proof has been found.

If the first reviews are to be believed, the model was otherwise a step forward. Anthropic says the answers are 40 percent less wordy than in its predecessor Opus 5 and that the model writes more than 30 percent faster. The newsletter Every has written that it produces "the most readable prose" of any model from Anthropic or OpenAI. The default setting has also been moved from "high" to "medium" effort, which means the subscription goes further.

A benchmark keeps watch

To test the rumor, a project called Livenerf has been started. It is a 30-day test where the method has been fixed in advance. Every day, the same 78 questions are put to the model via a Claude Max subscription. It measures how correct the answers are and how many words the model uses. Changes in answer length can be an early sign that something has changed, before correctness drops.

In an initial test, the creator found that a lower effort level significantly cuts answer length. He also found that there was hardly any difference between Opus 5.5 and Opus 5 in his test.

Skepticism in the debate

In the discussion on Hacker News, some say they have noticed a decline after the launch. Others think it is due to a "honeymoon effect": you are impressed at first and become more critical once the novelty has worn off or when you use the model for harder tasks.

Nor is it entirely out of the question that such complaints may have something to them. Another benchmark, Nerf Bench, previously detected a decline in Opus 4.6, which Anthropic acknowledged. Anthropic itself has said that the company does not lower quality because of demand or load, but in a review from September 2025 it admitted that three technical bugs degraded Claude's answers for a period.

What does it mean for you?

If you use Claude at work, there is no reason to panic. There is not yet any documented degradation of Opus 5.5. If you have an important task, save examples of good answers now so you can compare later. The results from Livenerf will give a more sober picture in about a month.