Anthropic launches Claude Fable 5.1 and restricted Mythos 5.1 with lower prices and research wins

One model, two safety postures

Anthropic has introduced Claude Fable 5.1 and Claude Mythos 5.1, which it calls its most advanced models yet for coding and knowledge work. According to the company's September 1 announcement, surfaced on Hacker News, the two are the same underlying model with different safeguards. Fable 5.1 is generally available, while Mythos 5.1 is restricted to trusted access programs tuned for cybersecurity and life sciences work.

Cheaper tokens and new data controls

Fable 5.1 should cost about 25% less than Fable 5 for typical workloads billed by token, a cut Anthropic attributes to lower pricing on cache reads. For heavily agentic work, savings can reach roughly 45%. The model also defaults to High effort in Claude Code and Medium effort in Claude Cowork and on Claude.ai.

On privacy, Anthropic is rolling out Enterprise Frontier Safeguards (EFS), a system it says provides privacy equivalent to zero data retention while remaining strong against adversarial use. Data is stored in cloud infrastructure controlled entirely by the customer rather than by Anthropic. EFS reaches enterprise customers in phases starting later this fall; until then, eligible customers can run Fable 5.1 under zero data retention.

Safeguards have also been tuned to cut false positives by 60% in cybersecurity. Fable 5.1 may now be used to find software vulnerabilities, though not to develop exploits for them. A separate access program for Mythos 5.1's advanced biology capabilities, built with the US government, is expected to open enrollment for scientists soon.

Benchmark results, with caveats

Anthropic says Fable 5.1 sets a new performance frontier. On Terminal-Bench-Science 0.1, an agentic scientific research benchmark, it scored 52.6%, against 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. On Terminal-Bench 4.0 agentic coding, the restricted Mythos 5.1 scored 60.9%, ahead of Fable 5.1's 55.8%. Elsewhere the gains are narrower: 1853 on GDPval-AA v2 versus 1824 for Opus 5, 73.4% on CursorBench 3.2.0, and 60.9% without tools on Humanity's Last Exam.

Anthropic flags its own caveats. Production safeguards intervened on some OSWorld 2.0 and AutomationBench tasks, routing cybersecurity work to Claude Opus 4.8 and biology tasks to Opus 5, which likely depresses those scores. Terminal-Bench-Science results carry a standard error of 3.5 to 4.5 points, and the OSWorld run used an August 2026 task release that is not comparable to earlier published numbers.

Customer anecdotes accompany the benchmarks. In testing by investment firm Millennium, Fable 5.1 reportedly found the cause of a rare crash in internal systems that the firm's engineers and other models had failed to explain over several years. Craig Falls, head of quantitative research at Jane Street Capital, said Fable 5.1 solves more of the firm's coding problems than Fable 5 or Opus 5, sets a company-best mark on trading intuition, and stays readable across long multi-step tasks where earlier models became harder to follow.

Science results take center stage

The most eye-catching claims involve research. Given open-source protein design and folding tools, Mythos 5.1 produced high-affinity protein binders that two external organizations confirmed in the lab. On three targets, binding affinities were 10 times higher than the best entries in Adaptyv Bio's protein design competitions, and its hit rate neared 50% across 12 targets, versus the 10-15% Anthropic says is typical in the field.

Fable 5.1 also trained a neural network to build a new elevation map covering a third of Venus from NASA Magellan radar imagery gathered more than 30 years ago. The map resolves features down to two to three kilometers instead of 10 to 20, and measures heights up to 25% more accurately. Anthropic released it under a Creative Commons license ahead of NASA's VERITAS and ESA's EnVision missions so those teams can pick observation targets.

In computational biology, Mythos 5.1 wrote custom GPU kernels and cached intermediate results, speeding up seven open-source deep learning models by up to 2.5 times with identical outputs, meaningful because researchers may run such models thousands of times per experiment.

Why it matters

The release signals that frontier AI competition now runs on enterprise terms as much as raw capability: lower token prices, retention-free privacy options and fewer false-positive refusals are becoming headline features. The twin-model approach, a broadly available variant plus a government-coordinated restricted one for sensitive domains, offers a template for how labs may commercialize risky capabilities. Above all, Anthropic presents the protein design, Venus mapping and GPU kernel results as early evidence that models are shifting from assisting scientists to contributing discoveries directly, with implications for drug development, planetary science and the compute costs of biology research.