On August 20, X account Token Gremlin (@TokenGremlin), which tracks the AI frontier, posted a compilation of user reports about Claude Opus 5 that drew 55,000 views. It says complaints keep turning up that benchmark scores are high but the real-world experience is worse than earlier models.
Why has Opus 5 gotten so strange?
The compiled reports commonly mention strange sentence structure and hard-to-understand phrasing, hallucinated facts that don't exist, repeatedly redoing the same task and fixing it on its own, ignoring instructions spelled out in CLAUDE.md, starting work that wasn't requested, wasting time fixing bugs it created itself, and reporting unfinished work as complete. Issues describing the same symptoms are also open in GitHub's anthropics/claude-code repository: #83691 "in-session error rate rising" and #82162 "Opus 5.0 nerfed".
August 17-19: degradation the status page acknowledged too
These reports aren't just baseless rumor. Anthropic's official status page (status.anthropic.com) logged an incident of degraded model output quality for Opus 5 and Haiku 4.5, and said that after identifying the cause it completed a fix on August 19. Separately, there was also an outage affecting Opus 5 between 16:11 and 18:23 UTC on August 18. That users' experience overlaps in time with the official incident records is evidence that these complaints aren't just imagination.
No evidence of intentional downgrading, but
Token Gremlin himself drew a line: "This degradation isn't evidence that it's the cause of the odd behavior, and I found no evidence that Anthropic intentionally downgraded the model." Still, if you use Claude Opus 5 for coding, it's worth getting in the habit, when you get a strange response, of checking status.claude.com and status.anthropic.com for a current incident before blaming your prompt. If the same symptoms keep recurring, adding your case to the GitHub issues can also speed up finding the cause.