Anthropic releases Claude Fable 5.1 and Mythos 5.1 at 25% lower cost

Summary

Anthropic has launched Claude Fable 5.1 (general access) and Mythos 5.1 (restricted to verified life science and cybersecurity organizations). The company claims that agentic scientific research scores and protein design success rates exceeding human team performance have nearly doubled. This development also brings a 25% price reduction, driven by cheaper cache reads.

Why it matters

The jump from 24.7% to 52.6% on Terminal-Bench-Science is seen as a qualitative leap in AI-assisted research rather than incremental improvement. This holds potential to shorten timelines in drug discovery and computational biology. Externally verified protein binder results (50% success rate, compared to the industry norm of 10-15%) could reshape the economics of early-stage R&D if they hold up in production conditions. The limited Mythos tier and new anti-distillation controls also indicate that frontier model distribution is becoming a regulated, access-controlled market.

Highlights

• The 25% cost reduction is actually a 75% cut in cache read prices. Savings in long-context agent workflows can reach up to 45%, but this figure largely depends on cache reuse rates.
• Mythos 5.1's protein designs showed binding affinities 10 times higher than the best human submissions across three Adaptyv Bio targets. These results were verified by two external laboratories.
• New anti-distillation rules prevent manual context manipulation in new API accounts. This effectively closes a documented technique used to extract Claude's internal reasoning.

Critical take

The protein design headline relies on only 12 targets and three competition comparisons. As the article itself acknowledges, we are still far from the 'decade of work' required between lab verification and clinical production. Additionally, the 25% price cut is not a universal discount but a change in cache read accounting. Therefore, actual savings vary significantly depending on the workload.