Claude Fable 5.1

Claude Fable 5.1 is the generally available deployment of the frontier model Anthropic released on September 1, 2026. It and Claude Mythos 5.1 are, in Anthropic's description, "the same model, but with different levels of safeguards": Fable 5.1 ships with safeguards that block certain tasks in dual-use biology and cybersecurity domains, while Mythos 5.1 runs the same weights under more permissive safeguards and is reachable only through trusted-access programs (Source: anthropic.com). The pair succeeds Claude Fable 5 and Claude Mythos 5, released in June 2026 under the same two-configuration design.

Capabilities and benchmarks

Anthropic reports Fable 5.1 outperforming Fable 5 and Opus 5 across coding, knowledge work, computer use and long-horizon agentic tasks, with the largest gains in terminal-based scientific and engineering work. Scores were obtained with Fable 5.1's production safeguards enabled; on tasks where those safeguards intervened, cybersecurity tasks were completed by Opus 4.8 and biology tasks by Opus 5, and both Fable 5.1 and Fable 5 scored zero on affected OSWorld 2.0 tasks (Fable 5 also on AutomationBench), which Anthropic says likely depresses the reported figures (Source: anthropic.com).

Anthropic gives the Terminal-Bench-Science standard error as ±3.5–4.5 points per model and notes that its harness reproduces the public leaderboard's Opus 5 (30.0%) and Fable 5 (21.4%) figures at 29.0% and 24.7%, within noise. OSWorld 2.0 scores are on the benchmark authors' August 2026 task release, with Fable 5 and Opus 5 re-run under the same conditions and no competitor score shown because the task files differ from earlier releases. The Terminal-Bench 4.0 gap between Fable 5.1 and Mythos 5.1 reflects tasks on which the earlier, less precise cyber safeguards intervened; Anthropic expects the gap to narrow under the revised safeguards. Fable 5.1 defaults to High effort in Claude Code and to Medium in Claude Cowork and on claude.ai.

On the Anthropic ECI, Anthropic's internal fork of Epoch AI's Epoch Capabilities Index, the model scores 161.98 (95% CI 158.20–169.00), against 160.73 for Opus 5 and 159.46 for Mythos 5 — a position Anthropic reads as consistent with the pre-Mythos-Preview frontier trend of about 13.5 ECI per year rather than a further slope change (System Card: Claude Fable 5.1 & Claude Mythos 5.1 (Anthropic, September 2026)).

Anthropic reports several scientific-research results obtained with the model. Given open-source protein design and folding tools, Mythos 5.1's binder designs achieved binding affinities ten times higher than the best submissions to Adaptyv Bio's protein design competitions on three targets (EGFR, Nipah G, 15-PGDH), with a hit rate near 50% across twelve targets against a typical 10–15%; a de novo entry targeting the Nipah G stalk from Nick Boyd and Escalante Bio reached a comparable ~1.4 nM. Fable 5.1 trained a neural network producing a new elevation map of a third of Venus from Magellan radar imagery, resolving features at two to three kilometres rather than ten to twenty, released on Zenodo under a Creative Commons licence ahead of the NASA VERITAS and ESA EnVision missions. Mythos 5.1 wrote custom GPU kernels that sped up seven open-source protein and genomics models by up to 2.5× with identical outputs, cutting estimated GPU cost on genome-wide analyses by 30–60% (Source: anthropic.com).

A later result in theoretical physics used the model inside the Claude Science harness. According to an account Anthropic published on September 25, 2026, written by physicist Matt von Hippel, Anthropic physicists used Fable 5.1 to compute the nine-loop MHV six-particle amplitude in planar N=4 super Yang-Mills theory, which Lance Dixon of SLAC and Stanford validated after being told of it on September 1; the bootstrap computation took about $100 of compute, equivalent to 96 CPUs for a week, within a total end-user cost of about $1,000 to $2,000 per method. Von Hippel concludes that the model used known methods (Source: anthropic.com). See AI for Science.

Safety and evaluations

The system card treats Mythos 5.1 as having CB-1 chemical and biological capabilities — the conservative resolution applied to previous models — and determines explicitly that it does not cross CB-2, citing weak novel ideation, poor strategic judgment, poor technical calibration and errors requiring expertise to catch. Autonomy risk from automated AI R&D is again assessed as low, with METR concurring that the model is "likely unable to fully and reliably automate R&D for frontier projects spanning multiple weeks" while "still likely to noticeably accelerate researchers." Catastrophic alignment risk is assessed as low — the level to which the August 2026 Risk Report had raised it from very low following incident disclosures about model behavior in cybersecurity evaluations — with no update to specific claims implying a further increase. Full detail is at System Card: Claude Fable 5.1 & Claude Mythos 5.1 (Anthropic, September 2026).

Anthropic states it has found no evidence of a critical-severity jailbreak for Fable 5.1's safeguards, having commissioned external robustness testing from two organizations plus automated testing by Gray Swan alongside its own dynamic evaluation (Source: anthropic.com).

Safeguards and access

Fable 5.1's cyber safeguards run in two stages: a probe over the model's internal activations screens all traffic and escalates anything flagged as cyber-related to a trained LLM classifier, which decides in combination with the probe's verdict whether to block. On most interfaces, flagged requests fall back to Opus 4.8; because the classifiers fire consistently across Anthropic's cyber evaluations, the company concludes Fable 5.1 "does not provide an uplift on cyber tasks relative to Opus 4.8" and publishes no separate Fable 5.1 cyber results (System Card: Claude Fable 5.1 & Claude Mythos 5.1 (Anthropic, September 2026)).

Two safeguard changes accompanied the release. Fable 5.1 may now be used to identify software vulnerabilities in source code at all access levels, though not to develop exploits for them; Anthropic reports the revised cyber safeguards blocking 60% fewer false positives and Claude Code users seeing around 60% fewer interventions per session than under Fable 5's safeguards. Penetration testing, exploit generation and binary-based vulnerability scanning continue to be redirected to Opus models. In biology, safeguards fire 85% less often for benign elementary-biology and medical questions than those that launched with Fable 5, while life-sciences research and development queries are still routed to Opus models (Source: anthropic.com).

Anthropic also introduced Enterprise Frontier Safeguards, which it describes as giving customers "the same as a zero data retention policy" by storing data in customer-controlled cloud infrastructure with human review performed by the customer by default. It was developed with more than 100 customers and with Amazon Web Services, Google Cloud and Microsoft Azure, and rolls out in phases from autumn 2026 across Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform and Microsoft Foundry; eligible customers may use Fable 5.1 and Fable 5 with zero data retention until it ships. This replaces the 30-day retention requirement that had applied to Mythos-class traffic (Source: anthropic.com).

A distillation countermeasure shipped with the release: API accounts created from September 1, 2026 onward can no longer manually edit Claude's prior context in a multi-turn conversation while preserving the transcript of its prior thinking. Anthropic describes this as closing "a common, publicly documented distillation technique," applies it gradually to existing accounts, and says it will apply to all users with future model releases (Source: anthropic.com).

Fable 5.1 and Mythos 5.1 carry a text watermark, which Anthropic describes as "a numerical way of determining the likelihood that Claude was involved in writing a piece of text." It follows from Anthropic's July 2026 signature — alongside 190 other signatories — of the EU AI Act's Code of Practice on Transparency of AI-Generated Content, which requires watermarking outputs of models released after August 2, 2026. Anthropic states the watermark is invisible without the detection API, has no effect on output quality or content, and carries no information about the user, their organization or their conversations. A detection API is in private preview, available to eligible organizations including regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, and to enterprises with their own compliance obligations under the Act (Source: anthropic.com; anthropic.com).

Mythos 5.1 access runs through two named programs. The Cyber Verification Program currently supplies Opus- and Sonnet-class models with reduced cyber safeguards for defensive security work, and Anthropic says it will include Mythos-class models "in the near future." The Life Sciences Verification Program supplies Mythos 5.1 with safeguards designed for professional research and development while other safeguards remain in place; Anthropic states it enrolled its first participants in partnership with the US government and plans to expand access to the broader life-sciences community (Source: anthropic.com).

Availability and pricing

Fable 5.1 is available on the Claude API as claude-fable-5-1 and on Amazon Web Services, Google Cloud and Microsoft Azure from launch. Input and output pricing is unchanged from Fable 5 at $10 and $50 per million tokens; cache reads fell 75% to $0.25 per million. Anthropic gives the resulting cost reduction as roughly 25% for typical workloads and up to roughly 45% for highly agentic, context-heavy work, measured at usage-based pricing at default effort over four weeks of August 2026 usage (Source: anthropic.com).

Reception

Anthropic published testimony from early-access partners alongside the launch. Cognition said it was moving Devin's Opus 5 traffic to Fable 5.1 on launch day, citing the cache-read pricing as making a Fable-class model economical for code review. Browserbase reported 82% task completion on its hardest browser-agent benchmark at roughly ten minutes per task, against 74% for Opus 5 and 57% for Fable 5, using fewer tokens than either. Crosby reported its RedlineBench contract-redlining score rising from 47.9 to 57.0, with most of the gain in first-turn quality. Samaya reported a 55.9% rubric score on its FrontierFinance benchmark against 49.2% for Fable 5. Rogo reported matching Fable 5's accuracy on its internal finance benchmark with 20% fewer tokens. Millennium described the model diagnosing a one-in-a-million crash that had gone unexplained for four to five years by disassembling a vendor library and matching it against a core dump (Source: anthropic.com). These are vendor-published customer statements rather than independent evaluations.

Relationships

• instance-of: Frontier Models
• supersedes: Claude Fable 5
• related: Claude Mythos 5.1 — the same model under more permissive safeguards
• depends-on: System Card: Claude Fable 5.1 & Claude Mythos 5.1 (Anthropic, September 2026)
• related: Anthropic, Responsible Scaling Policy (RSP), Frontier Compliance Framework (February 2026), CBRN Uplift