On Tuesday, Anthropic launched Fable and Mythos 5.1, twinned variations of the corporate’s most superior AI mannequin. Along with efficiency upgrades, the brand new Fable launch consists of modifications meant to cut back token price and false-positive restrictions from the mannequin’s safeguards.
As with the earlier Mythos mannequin, Mythos 5.1 will solely be obtainable to registered Anthropic companions engaged in both cybersecurity or life sciences analysis. Fable 5.1, the unrestricted model, is on the market beginning immediately on cloud platforms or via the Anthropic API.
One of the important modifications is Anthropic’s previously reported embrace of zero knowledge retention, permitting shoppers to run Anthropic fashions on their very own infrastructure with out knowledge outflows. Beforehand unavailable for Fable because of safety considerations, a high-privacy service (referred to as Enterprise Frontier Safeguards) will now roll out to customers within the fall. Notably, the system will nonetheless monitor for misuse by brokers or human customers, however shoppers will management how the monitoring takes place.
As a part of the announcement, Anthropic assured clients that their knowledge had not been inappropriately accessed. “Anthropic has by no means skilled on enterprise knowledge with out express permission, and by no means will,” the announcement reads.
As is frequent for an Anthropic launch, the brand new fashions set information in a spread of benchmarks, together with Terminal-Bench 4.0 (for CLI-based coding) and Humanity’s Final Examination (for basic reasoning). Anthropic additionally launched three novel scientific findings generated by the fashions earlier than their launch, together with a customized GPU optimization and a high-resolution map of Venus assembled from current pictures.
As with earlier releases, the fashions include a detailed system card, which explains their capabilities in most easy phrases. The system card charges Mythos as “low-risk” for considerations associated to automated AI improvement — the place the AI improves itself — which some see as a set off for a lack of human management. It says “its skill to speed up inside AI R&D progress is in keeping with present tendencies.”
When it comes to basic misbehavior, Mythos is barely extra vulnerable to it than Opus, probably on account of its enhanced capabilities.
“Mythos 5.1 is a slight regression on total misaligned habits in comparison with Opus 5, and an enchancment over Mythos 5 and Claude Sonnet 5,” the system card reads. “It cooperates with human misuse and accepts unverifiable claims of authorization considerably extra readily than Opus 5, however it’s much less more likely to ignore express constraints, hallucinate inputs, or falsely declare to have accomplished duties than earlier fashions.”
Whenever you buy via hyperlinks in our articles, we could earn a small fee. This doesn’t have an effect on our editorial independence.