Fable 5.1

Model Drop: Claude Fable 5.1

Claude Fable 5.1 launched today alongside its unmuzzled twin Claude Mythos 5.1. Same weights, two sets of safeguards, one static price, and a cache-read discount that shows Anthropic is watching OpenAI very closely.

What shipped

Pricing: $10 / $50 per million input / output tokens, identical to Fable 5 and double Opus 5’s $5 / $25. Cache reads are now $0.25 per million, down from $1 on Fable 5. Cache writes ($12.50 five-minute, $20 one-hour) and the 512-token minimum are unchanged. Batch is $5 / $25.

Available on: claude.ai and the Claude apps on Pro, Max, Team, and Enterprise (Free is not mentioned anywhere), Claude Code, Claude Cowork, the Claude Developer Platform, Amazon Bedrock and Claude Platform on AWS, Google Cloud including the Agent Platform, and Microsoft Foundry. Mythos 5.1 is reachable only through the Cyber Verification Program, the new Life Sciences Verification Program, and Project Glasswing account teams. Its capabilities also power Claude Security, the codebase-scanning product every Claude Enterprise customer gets.

Other info: 1M-token context window (default and maximum), 128K max output. Reliable knowledge cutoff June 2026, the freshest of any Claude model. Same tokenizer as everything since Opus 4.7 (roughly 30% more tokens than pre-4.7 models for the same text). Under the Responsible Scaling Policy, Anthropic judges the model at CB-1 for chem and bio (”could meaningfully help someone with a basic technical background synthesize a known weapon”) but short of the CB-2 threshold, holds that judgment “with some uncertainty,” and has moved its overall alignment-risk assessment from “very low” to “low.” Strongest cyber capabilities Anthropic has ever released, still inside the lower risk tier of its Frontier Compliance Framework. Safeguard changes: source-code vulnerability discovery is allowed at general availability (as on Opus 5), cyber classifiers intervene about 60% less per session than Fable 5’s, bio classifiers throw 85% fewer false positives on basic biology and medical questions; penetration testing, exploit generation, and binary vulnerability scanning still route to Opus. 212-page system card published at launch. External red teaming by Trajectory Labs: 74 hours, 6,500-plus requests, no working end-to-end exploit and no universal jailbreak.

More details: Introducing Claude Fable 5.1 and Claude Mythos 5.1, the system card, and What’s new in Claude Fable 5.1 for developers.

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 this morning, twelve weeks after Fable 5 and five weeks after Opus 5, and pitched the pair as “the world’s most advanced models for coding and knowledge work” with “an early glimpse of how AI models will contribute to scientific progress.” The structure is the one Anthropic invented in June: one set of weights, two products. Fable 5.1 is generally available on every platform Anthropic sells through and on the Pro, Max, Team, and Enterprise plans, with safeguards that hand cyber and bio requests to an Opus model. Mythos 5.1 is the same model with those safeguards relaxed, and the audience for it has grown from Project Glasswing’s cyberdefenders to a second trusted-access track for life-science researchers, run with the US government and currently limited to American organizations. The price stayed at $10 / $50, and the one line that did change is cache reads, cut 75% to 25 cents per million tokens.

The evidence is broader than the Fable 5 launch and, mercifully, comes with actual numbers instead of the ratios Opus 5 hid behind. Terminal-Bench-Science more than doubles Fable 5. AutomationBench nearly doubles it. CursorBench, Humanity’s Last Exam, OSWorld, and GDPval all move in the right direction by small, credible margins. Protein binders with affinities ten times better than the top submissions to Adaptyv Bio’s design competitions and a near-50% hit rate across twelve targets, a new elevation map of a third of Venus at two-to-three-kilometer resolution released under Creative Commons on Zenodo, and seven genomics models sped up 1.4x to 2.5x with 30 to 60% GPU cost savings in a few days of unattended work.

Anthropic admits Fable 5 “scored zero” on AutomationBench tasks where the safeguards fired. The system card moves alignment risk from very low to low, reports a sandbox escape during external testing, and describes a model that is “less honest under pressure than recent Claude models” and “among the most capable models we have tested at controlling the contents of its extended thinking.” And the gap between Mythos 5.1’s 60.9% and Fable 5.1’s 55.8% on Terminal-Bench 4.0 is, per Anthropic, entirely the safeguards intervening.

What’s new

This is a point release, so most of the novelty is in the API contract and the deployment policy rather than true differences in the weights.

The benchmarks

Anthropic’s numbers

• Terminal-Bench-Science 0.1: 52.6%. +27.9 over Fable 5 (24.7%), +23.6 over Opus 5 (29.0%), +30.2 over GPT-5.6 Sol (22.4%).

• Terminal-Bench 4.0: 60.9% (Mythos 5.1), 55.8% (Fable 5.1)

  • +18.9 (Mythos 5), +23.6 (GPT-5.6 Sol)

• GDPval-AA v2 (knowledge work): 1853. +29 over Opus 5 (1824), +142 over Sol (1711).

• Humanity’s Last Exam without tools: 60.9%. +3.1 over Fable 5 (57.8%), +4.3 over Opus 5 (56.6%).