Claude Opus 5 is fast, cheap, and strong — but still not the top-tier Mythos model

Don't Worry About the Vase (Zvi) · rss · 2026-07-29

This long review argues that Claude Opus 5 is a very capable model, but not quite in the same tier as the reviewer’s top “Mythos” model.

Anthropic’s pitch is that Opus 5 matches most of Fable 5’s performance at about half the API price, with more permissive classifiers.
The reviewer says Opus 5 is roughly as capable as Fable 5 on most real-world tasks, and sometimes modestly better.
But it is still described as weaker than Mythos/Fable on the hardest, most agentic work, especially when it has to run the show rather than act as a subagent.
A recurring theme is “vibe”: some users reportedly dislike Opus 5’s verbosity, repetition, negativity, or confrontational tone.
The review highlights strong benchmark results, including claims of new SOTA on coding/knowledge-work evaluations, a perfect 42/42 on the 2026 IMO, and very strong performance in tool-using settings.
One especially notable claim is that Opus 5 is Anthropic’s least prompt-injectable model yet; layered defenses reportedly push prompt-injection success rates to near zero.
The reviewer’s practical conclusion: Opus 5 is a strong everyday model and a very good subagent, but they would still choose Fable 5 when they can.

More from Models