AI August 14, 2026 bearish ⇧ 182 pts across 1 thread

Claude Opus 5 Quality Complaints Mount

A thread asking 'Why does Opus 5 feel worse to work with?' surfaced real frustration from active users. Complaints included degraded code output quality compared to 4.5, slower completion times, and outputs that are hard to parse or act on. One commenter shared an example where the model returned fragmented, nearly incoherent instructions. Another noted that for non-English speakers (specifically Italian), the model 'approximates' words in ways that introduce errors.

This is notable because Anthropic has positioned Opus as its premium, high-reasoning tier. If the flagship is regressing in perceived quality while competitors are shipping at lower price points, that's a real retention problem. The complaints aren't about a niche use case either: they're about everyday coding and writing tasks.

The counterpoint is that model quality perception is subjective and prompt-sensitive. But when multiple commenters with different use cases and languages converge on the same 'it got worse' conclusion, that's not just vibes.


So what?

If you're paying Opus-tier prices and routing your most complex tasks there, run your own evals before assuming the latest version is better than the previous one. Model versioning is not the same as software versioning and 'latest' does not mean 'best for your use case'.

Read these