On Friday, Anthropic launched the Opus 5 model, the latest version of its long-standing heavyweight model. Although this model is smaller than Fable 5, it is cheaper and less restrictive than Fable, so we believe it is preferred for most use cases.
Notably, Opus 5 actually outperforms Fable 5 on many of the benchmarks included in the announcement.
Opus 5 is released just two months after Opus 4.8, which became available on May 28th. Mythos 5, Fable 5, and Sonnet 5 were all released in June, and only the lightweight Haiku model is still waiting to upgrade to the 5 series.
In a post announcing the new model, Anthropic highlighted how Opus 5 is “much better at validating behavior and carefully iterating until it succeeds,” citing examples such as benchmark tests where Opus 5 responded to incomplete prompts and created its own computer vision pipeline.
Importantly, Opus 5 is also free from many of the limitations that have dogged Fable since its release. Like its predecessor, Opus 5 is not subject to the 30-day data retention policy covering Fable and Mythos, which has raised concerns among some privacy-minded users.
Opus still has meaningful safeguards, especially when it comes to cybersecurity tasks such as exploit generation and penetration testing. For example, Opus 5’s safety measures prohibit its use to scan software binaries for vulnerabilities, but allow it to search for vulnerabilities in source code, as the latter task is likely to be used for defensive purposes.
Roughly speaking, Anthropic expects these classifiers to be used 85% less frequently in Opus 5 than in Fable 5. This reflects a lighter touch being given to less capable models.
Anthropic is also rolling out new tools to reduce disruption when safety equipment is activated. Users can now select a beta feature called automatic fallback. This automatically routes requests to a less powerful model when a prompt triggers the safety classifier. As a result, API users who enable the setting will receive a functional response instead of an error message.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
