Anthropic has just released Opus 5, its latest flagship AI model. It is cheaper than Fable 5 and has fewer restrictions, meaning it might be the more suitable option for the average user. Although it is smaller in size, Opus 5 outperforms Fable 5 on a number of benchmarks that Anthropic provided in its announcement.
Anthropic Keeps Up a Fast Release Pace
Anthropic is not stopping. The company launched Opus 4.8 just two months ago, on May 28, and in June issued Mythos 5, Fable 5, and Sonnet 5. The only model that hasn’t had its version 5 update is Haiku.
The Model Checks Its Own Work
Anthropic says Opus 5 handles complex tasks better than before. The model verifies its own work and keeps refining it until it succeeds. In one test, Opus 5 built a full computer vision pipeline from a prompt that left out key details.
Fewer Guardrails, But Not None
Opus 5 skips a rule that bothers privacy-conscious users. Anthropic doesn’t apply the 30-day data retention policy to Opus 5, the same as it did for Opus 4.8, even though this rule still covers Fable and Mythos.
Some limits remain, though. Anthropic still restricts cybersecurity tasks like exploit generation and penetration testing. For example, Opus 5 can’t scan a software binary for vulnerabilities, but it can search for vulnerabilities in source code instead, since that work usually serves defensive purposes.
Overall, the model faces far fewer blocks. Anthropic expects its safety filters to trigger 85% less often on Opus 5 than on Fable 5. That’s a lighter touch for a model that ranks below Fable in raw capability.
A Smoother Fallback When Filters Trigger
Anthropic also rolled out a new tool to soften the impact of its safety filters. Users can now turn on Automatic Fallbacks, a beta feature that sends a blocked request to a weaker model instead of rejecting it outright. API users who enable this setting will receive a real answer instead of an error.
