On Friday, Anthropic introduced its latest model, Opus 5, which is part of the company’s series of advanced models. Although it is smaller than Fable 5, Opus 5 is designed to be more affordable and less restrictive, making it a more attractive choice for many applications.
Significantly, Opus 5 surpasses Fable 5 in various benchmarks highlighted in the announcement.
The launch of Opus 5 comes just two months after Opus 4.8, which was released on May 28. Other models, including Mythos 5, Fable 5, and Sonnet 5, were all introduced in June, leaving only the lighter Haiku model awaiting an update to the 5 series.
In its announcement, Anthropic underscored that Opus 5 is remarkably enhanced in its ability to validate its output and improve through careful iterations. Notably, benchmark tests demonstrated Opus 5’s capability to autonomously develop a computer vision pipeline in response to incomplete prompts, among other achievements.
Importantly, Opus 5 eliminates many of the limitations that have affected Fable since its debut. Like its earlier version, Opus 5 is not bound by the 30-day data retention policy that applies to Fable and Mythos, addressing concerns from privacy-minded users.
Nonetheless, Opus retains significant safeguards, especially concerning cybersecurity functions like exploit generation and penetration testing. For example, while Opus 5 is restricted from scanning software binaries for vulnerabilities, it is permitted to examine source code for vulnerabilities, as that task is generally more aligned with defensive strategies.
In general, Anthropic anticipates that the safety classifiers will engage 85% less frequently with Opus 5 compared to Fable 5, illustrating a more lenient approach towards the less powerful model.
Additionally, Anthropic is launching a new feature aimed at minimizing disruption from safety interventions. Users can now opt-in to a beta feature known as Automatic Fallbacks, which automatically redirects requests to a less powerful model if a prompt activates the safety classifier. This means that API users who enable this setting will receive a functional response rather than an error message.
