Anthropic has launched Opus 5, a new version of its flagship model that outperforms Fable 5 on several benchmarks despite being smaller. The release follows a rapid schedule with Opus 4.8 arriving in May and Mythos 5, Fable 5 and Sonnet 5 launching in June, leaving only the Haiku model without a version 5 upgrade. This latest iteration focuses on verification and careful iteration, with examples including writing a computer vision pipeline from an incomplete prompt. Crucially, Opus 5 removes the 30-day data retention policy that previously concerned privacy-conscious users regarding Fable and Mythos. Safeguards remain active for cybersecurity tasks like exploit generation, though the model can now search for vulnerabilities in source code for defensive purposes. Anthropic expects these safety classifiers to engage 85% less often than with Fable 5. A new Automatic Fallbacks beta feature also allows users to route blocked requests to a less powerful model, returning a functional response instead of an error.
- Opus 5 is cheaper and less restrictive than the larger Fable model.
- Data retention for Opus 5 does not extend for 30 days.
- Automatic Fallbacks prevents API errors when safety filters trigger.
