Anthropic launches Opus 5 with fewer restrictions
On Friday, Anthropic unveiled Opus 5, its newest robust model, which surpasses Fable 5 in various benchmarks while being more affordable and less restrictive. This release follows just two months after Opus 4.8 on May 28. In June, Mythos 5, Fable 5, and Sonnet 5 were introduced, leaving only the lighter Haiku model pending a 5-series upgrade. Anthropic emphasized Opus 5's enhanced self-verification and iterative problem-solving capabilities, notably its proficiency in creating a computer vision pipeline from a partial prompt. Unlike Fable, Opus 5 does not adhere to a 30-day data retention policy, alleviating privacy issues. Nonetheless, it maintains safeguards for cybersecurity tasks, such as exploit generation and penetration testing, with limitations on scanning software binaries for vulnerabilities. Anthropic anticipates that Opus 5's classifiers will engage 85% less frequently than those of Fable 5. A new beta feature, Automatic Fallbacks, enables API users to redirect requests to a less powerful model when safety classifiers are triggered, ensuring a functional response instead of an error.
Key facts
- Anthropic launched Opus 5 on Friday.
- Opus 5 outperforms Fable 5 on several benchmarks.
- Opus 5 is cheaper and less restrictive than Fable 5.
- Opus 5 launched two months after Opus 4.8 (May 28).
- Mythos 5, Fable 5, and Sonnet 5 launched in June.
- Haiku model still awaiting 5-series upgrade.
- Opus 5 is not subject to 30-day data retention policy.
- Safeguards prevent Opus 5 from scanning software binaries for vulnerabilities.
- Opus 5 can search for vulnerabilities in source code for defensive purposes.
- Classifiers engage 85% less often for Opus 5 than Fable 5.
- New beta feature Automatic Fallbacks routes requests to less powerful model when safety classifier triggers.
Entities
Institutions
- Anthropic
- OpenAI