On Friday, Anthropic released its Opus 5 model, the newest version of its long-standing heavyweight model. While smaller than the Fable 5, the model will be cheaper and less restrictive than the Fable, which will likely make it preferable in most use cases.

Notably, Opus 5 actually outperforms Fable 5 on a number of benchmarks included in the announcement.

Opus 5 launches just two months after Opus 4.8, which became available on May 28. Mythos 5, Fable 5, and Sonnet 5 all launched in June, leaving only the lightweight Haiku model awaiting an upgrade to the 5 series.

In a post announcing the new model, Anthropic emphasized that Opus 5 was “much stronger at verifying its work and carefully iterating until success,” citing benchmark tests, in which Opus 5 wrote its own computer vision process in response to an incomplete message, among other examples.

Crucially, Opus 5 is also free of many of the restrictions that have plagued Fable since its release. Like its predecessor, Opus 5 is not subject to the 30-day data retention policy that covers Fable and Mythos, which has raised concerns among some privacy-conscious users.

There are still significant safeguards in Opus, particularly around cybersecurity tasks such as exploit generation and penetration testing. For example, Opus 5 safeguards prevent it from being used to search for vulnerabilities in a software binary, although searching for vulnerabilities in the source code is permitted, as the latter task is more likely to be used for defensive purposes.

Generally speaking, Anthropic expects these classifiers to interact 85% less frequently for Opus 5 than for Fable 5, a reflection of the lighter touch given to the less capable model.

Anthropic is also implementing a new tool to make safeguards less disruptive when activated. Users can now opt for a beta feature called Automatic Reservations, which will automatically route requests to a less powerful model when a message triggers the security classifier. The result is that API users with the setting enabled will get a functional response instead of an error message.