On Friday, Anthropic unveiled Opus 5, the latest iteration of its flagship heavyweight model. Although smaller than Fable 5, Opus 5 is both more affordable and less restrictive, making it a more attractive option for most use cases. The model surpasses Fable 5 on several key benchmarks highlighted in the announcement.
Performance and Availability
Opus 5 arrives just two months after Opus 4.8, which became available on May 28, 2026. In contrast, Anthropic's other models—Mythos 5, Fable 5, and Sonnet 5—all launched in June, leaving only the lightweight Haiku model awaiting its 5-series upgrade.
In a blog post, Anthropic emphasized that Opus 5 is "much stronger at verifying its work and iterating carefully until it succeeds." The company cited benchmark tests where Opus 5 autonomously built a computer vision pipeline from an incomplete prompt, among other examples of its enhanced reasoning and self-correction capabilities.
Fewer Restrictions, Enhanced Privacy
A crucial advantage of Opus 5 is its freedom from many of the constraints that have affected Fable 5 since its release. Like its predecessor, Opus 5 is not subject to the 30-day data retention policy that applies to Fable and Mythos, addressing privacy concerns among users who prioritize data minimization.
Safeguards and New Features
Despite its lighter restrictions, Opus 5 retains meaningful safeguards, particularly around cybersecurity tasks such as exploit generation and penetration testing. For example, the model is prevented from scanning for vulnerabilities in software binaries, though it is allowed to search for vulnerabilities in source code—a task more commonly associated with defensive security efforts.
Anthropic expects these safety classifiers to engage 85% less often for Opus 5 compared to Fable 5, reflecting the lighter touch applied to this less capable but more versatile model.
To further reduce disruption when safeguards do trigger, Anthropic is introducing a new tool called Automatic Fallbacks. Users can opt into this beta feature, which automatically routes requests to a less powerful model when the safety classifier is activated. This ensures that API users receive a functional response instead of an error message, as noted in a related TechCrunch article on AI guardrails.
via TechCrunch AI