Anthropic's new Opus 5 beats its flagship Fable 5 on benchmarks and costs less

Anthropic launched Opus 5 on Friday, and the new model pulls off a strange trick: it’s cheaper and less restrictive than the company’s current flagship, Fable 5, yet it actually outperforms it on a number of benchmarks. For anyone who found Fable 5’s guardrails to be a drag on real work, this is a pretty clear signal that Anthropic is listening. The model arrives just two months after Opus 4.8, keeping pace with a frantic release cycle that saw Mythos 5, Fable 5, and Sonnet 5 all land in June.

The performance claims are worth paying attention to. Anthropic says Opus 5 is “much stronger at verifying its work and iterating carefully until it succeeds,” and they backed that up with a demo where the model wrote its own computer vision pipeline from an incomplete prompt. That kind of autonomous, multi-step engineering is the difference between a chatbot and a junior developer.

Privacy is a big part of the pitch here. Unlike Fable 5 and Mythos 5, Opus 5 isn’t subject to that 30-day data retention policy that made some enterprise users sweat. There are still meaningful safeguards—it won’t help you scan a software binary for vulnerabilities, though it will help with source code since that’s more likely to be defensive work. Anthropic expects its safety classifiers will trip 85% less often for Opus 5 than for Fable 5, which should mean fewer “I can’t help with that” dead ends.

Anthropic is also rolling out a beta feature called Automatic Fallbacks that aims to make the remaining safeguards less annoying. If a prompt trips the safety classifier, the system will automatically route the request to a less powerful model instead of just throwing an error. For API users, that’s the difference between getting a functional response and getting a brick wall. It’s a small quality-of-life change, but it reveals a philosophy shift: safety doesn’t have to mean refusal.