# anthropic-claude-fable-5-1-mythos-5-1-2026-09-01

## Veille

Product communication from **Anthropic** published on **September 1, 2026** on anthropic.com (~4,000 words, six sections, 22 testimonials from early-access partners). It announces **Claude Fable 5.1** (general availability) and **Claude Mythos 5.1** (verified access): *the same model, but with different levels of safeguards*.

## Titre Article

Claude Fable 5.1 and Mythos 5.1

## Date

2026-09-01

## URL

https://www.anthropic.com/claude-fable-and-mythos-5-1

## Keywords

Claude Fable 5.1, Claude Mythos 5.1, foundation model, cache reads, cache pricing, agentic cost, effort levels, Terminal-Bench-Science, Terminal-Bench 4.0, GDPval-AA, OSWorld 2.0, CursorBench, AutomationBench, Humanity's Last Exam, protein design, high-affinity binders, Adaptyv Bio, carte altimétrique de Vénus, Magellan, GPU kernels, genomics, Evo 2, Enterprise Frontier Safeguards, zero data retention, cyber safeguards, false positives, vulnerability identification, anti-distillation, Cyber Verification Program, Life Sciences Verification Program, EU AI Act, watermark, detection API, Responsible Scaling Policy, Frontier Compliance Framework

## Authors

Anthropic — communication produit publiée sur anthropic.com, sans signature individuelle.

## Ton

Profile: corporate communication from an AI lab in the first-person plural, a disciplined technical-promotional register, aimed at developers, technical decision-makers, and researchers. Three traits distinguish this text from a standard product announcement. First, the **explicit response to criticism**: the introduction lists pricing, data retention, and safeguards as customer feedback the release responds to, which frames the announcement as part of a conversation rather than a proclamation. Second, **metrological discipline**: scores are given as cost/accuracy curves across five effort levels rather than as isolated points, standard errors are published (±3.5-4.5 pts on Terminal-Bench-Science), gaps with the public leaderboard are flagged and attributed to noise, and a note specifies that production safeguards caused the model to score zero on certain tasks — meaning the published figures underestimate the model. Finally, **self-limitation**: the alignment section lists what the audit does not cover (very-long-context, multi-agent, impossible tasks) and acknowledges that the model can still circumvent approvals and classifiers. The 22 partner testimonials carry the promotional weight, each anchored to a named, quantified internal benchmark. The dominant rhetoric is not raw power but **precision**: finer-grained safeguards, better-targeted costs, more tightly bounded measurements.

## Pense-betes

- **Fable 5.1 and Mythos 5.1 are the same model**: only the safeguards differ. Fable is in general availability; Mythos goes through verified-access programs (CVP for cyber defense, LSVP for life sciences, the latter set up with the US government).
- **The price cut only affects the cache**: cache reads **−75%** to **$0.25/M tokens**. Input (**$10/M**) and output (**$50/M**) are unchanged. The actual gain therefore depends on the cache's share of the workload — **~25%** for typical use, **~45%** for heavily agentic use. Cache-free usage gains nothing.
- **Default effort levels**: `high` in [[claxton-anthropic-ai-native-sdlc-playbook-2026-08-21]] and Claude Code, `medium` in Claude Cowork and on claude.ai. At `low`/`medium` effort, Fable 5.1 matches or exceeds Fable 5 at a markedly lower cost — the most direct cost lever.
- **Benchmarks**: Terminal-Bench-Science 0.1 **52.6%** (vs 24.7% Fable 5, 29.0% Opus 5, 22.4% GPT-5.6 Sol); Terminal-Bench 4.0 **55.8%** Fable / **60.9%** Mythos; GDPval-AA v2 **1853**; CursorBench 3.2.0 **73.4%**; AutomationBench **31.4%** (vs 17.1%); HLE **60.9%** without tools.
- **The Fable/Mythos gap on Terminal-Bench measures the safeguards, not the capability** — Anthropic states that the new classifiers should reduce it significantly.
- **Enterprise Frontier Safeguards (EFS)**: data stays on the customer's own cloud infrastructure, and human review is performed by the customer. Phased rollout starting fall 2026 across Claude Code, Claude Enterprise, the platform, Bedrock, Google Agent Platform, and Microsoft Foundry. In the meantime, zero retention for eligible customers. Built with 100+ customers and the three hyperscalers.
- **Cyber opening**: Fable 5.1 can identify vulnerabilities, not generate exploits. Still redirected to Opus: penetration testing, exploit generation, binary vulnerability scanning. Biology safeguards trigger **85% less** on benign medical questions.
- **Anti-distillation**: API accounts created from September 1, 2026 onward can no longer manually edit Claude's prior context while retaining its reasoning trace. Existing accounts are spared for now, but the rule will apply to everyone in upcoming releases — worth checking if a custom integration depends on this mechanism.
- **EU AI Act compliance**: digital watermark on the outputs of models released after August 2, 2026, invisible without the detection API, with no information about the user. Detection API in private preview for regulators, media, researchers, and companies subject to the same obligation.
- **Science**: hit rate **~50%** across 12 protein targets (10-15% usual), affinities **10×** better on three targets (EGFR, Nipah G, 15-PGDH); carte de Vénus published under Creative Commons ahead of the NASA VERITAS and ESA EnVision missions; GPU optimizations announced as soon to be open source.
- **Acknowledged limitations of the alignment audit**: limited visibility into very-long-context and multi-agent work, insufficient coverage of impossible tasks — precisely the terrain documented by [[mollick-agency-and-agents-twilight-factory-2026-08-31]].

## RésuméDe400mots

On **September 1, 2026**, Anthropic announces **Claude Fable 5.1** and **Claude Mythos 5.1**, presented as the most advanced models for coding and knowledge work. Both are **the same underlying model**; only the levels of safeguards differ. Fable 5.1 is in general availability; Mythos 5.1 is accessible only through trusted-access programs, with safeguards designed for cybersecurity and the life sciences.

The announcement explicitly responds to three pieces of customer feedback. **Pricing**: cache reads drop by 75% to $0.25 per million tokens, with input and output unchanged at $10 and $50; total cost falls by around 25% on typical workloads and up to 45% on heavily agentic workloads. **Data retention**: the new *Enterprise Frontier Safeguards* store data on the customer's own infrastructure, offering the privacy of a zero-retention agreement while preserving detection of adversarial usage; phased rollout starting in the fall. **Safeguards**: cyber classifiers produce 60% fewer false positives, and Fable 5.1 is now permitted to identify software vulnerabilities — without developing exploits.

On performance, Fable 5.1 reaches 52.6% on Terminal-Bench-Science 0.1 (versus 24.7% for Fable 5 and 29.0% for Opus 5), 55.8% on Terminal-Bench 4.0 (60.9% for Mythos 5.1), 1853 on GDPval-AA v2, 73.4% on CursorBench 3.2.0, and 31.4% on AutomationBench. Results are presented as cost/accuracy curves across five effort levels; at low or medium effort, the model matches or exceeds Fable 5 at a much lower cost. Twenty-two partners give testimonials, including Millennium, where the model diagnosed a one-in-a-million crash that no one had explained in four to five years.

The science section documents three results. In **molecular design**, Mythos 5.1 achieves a success rate of nearly 50% across 12 protein targets, with affinities ten times higher than the best submissions from Adaptyv Bio. In **modeling**, Fable 5.1 produced a carte altimétrique de Vénus covering one third of Venus from Magellan radar data, published under a Creative Commons license. In **computational biology**, Mythos 5.1 accelerated seven open-source models by up to 2.5× by writing GPU kernels, cutting costs by 30 to 60%.

On safety, Mythos 5.1 stays below the next risk threshold of the Responsible Scaling Policy in biology and in the lower category of the Frontier Compliance Framework in cyber. The alignment audit finds it better aligned than Mythos 5, while acknowledging limited coverage of long-context, multi-agent, and impossible tasks.

## GrapheDeConnaissance

- Anthropic —publie→ Claude Fable 5.1 (TECHNOLOGIE, 0.99)
- Anthropic —publie→ Claude Mythos 5.1 (TECHNOLOGIE, 0.99)
- Claude Fable 5.1 —est_variante_de→ Claude Mythos 5.1 (TECHNOLOGIE, 0.97)
- Claude Fable 5.1 —remplace→ Claude Fable 5 (TECHNOLOGIE, 0.95)
- Claude Mythos 5.1 —remplace→ Claude Mythos 5 (TECHNOLOGIE, 0.95)
- Claude Fable 5.1 —utilise→ Enterprise Frontier Safeguards (TECHNOLOGIE, 0.92)
- Enterprise Frontier Safeguards —permet→ zéro rétention de données (CONCEPT, 0.93)
- Claude Fable 5.1 —réduit→ coût des lectures de cache de 75 %, à 0,25 $/M tokens (MESURE, 0.96)
- Claude Fable 5.1 —réduit→ coût total de ~25 % en charge typique et ~45 % en charge agentique (MESURE, 0.94)
- Claude Fable 5.1 —mesure→ 52,6 % sur Terminal-Bench-Science 0.1 (MESURE, 0.95)
- Claude Fable 5.1 —mesure→ 73,4 % sur CursorBench 3.2.0 (MESURE, 0.93)
- Claude Mythos 5.1 —mesure→ 60,9 % sur Terminal-Bench 4.0 (MESURE, 0.93)
- Claude Fable 5.1 —surpasse→ Claude Opus 5 (TECHNOLOGIE, 0.92)
- Claude Fable 5.1 —surpasse→ GPT-5.6 Sol (TECHNOLOGIE, 0.9)
- Claude Fable 5.1 —permet→ identification de vulnérabilités logicielles (CONCEPT, 0.93)
- garde-fous cyber —réduit→ 60 % d'interventions en moins par session dans Claude Code (MESURE, 0.92)
- Claude Mythos 5.1 —mesure→ taux de réussite de ~50 % sur 12 cibles de design de binders (MESURE, 0.9)
- Claude Mythos 5.1 —surpasse→ soumissions des compétitions Adaptyv Bio (DOCUMENT, 0.88)
- Claude Fable 5.1 —a_créé→ carte altimétrique de Vénus (DOCUMENT, 0.91)
- carte altimétrique de Vénus —est_basé_sur→ mission Magellan (EVENEMENT, 0.9)
- Claude Mythos 5.1 —améliore→ accélération jusqu'à 2,5× de sept modèles de deep learning (MESURE, 0.9)
- Claude Mythos 5.1 —réduit→ coûts GPU de 30-60 % sur analyses pangénomiques (MESURE, 0.88)
- Claude Mythos 5.1 —fait_partie_de→ Cyber Verification Program (CONCEPT, 0.88)
- Claude Mythos 5.1 —fait_partie_de→ Life Sciences Verification Program (CONCEPT, 0.9)
- Anthropic —collabore_avec→ gouvernement américain (ORGANISATION, 0.9)
- Claude Security —utilise→ Claude Mythos 5.1 (TECHNOLOGIE, 0.9)
- Anthropic —affirme_que→ Mythos 5.1 reste sous le palier de risque suivant de la Responsible Scaling Policy (AFFIRMATION, 0.92)
- Anthropic —affirme_que→ l'audit d'alignement couvre mal le long contexte, le multi-agent et les tâches impossibles (AFFIRMATION, 0.9)
- Claude Fable 5.1 —utilise→ mécanismes anti-distillation (TECHNOLOGIE, 0.91)
- Anthropic —utilise→ watermark de contenu généré (TECHNOLOGIE, 0.9)
- watermark de contenu généré —s_applique_à→ EU AI Act (DOCUMENT, 0.92)
- Millennium —affirme_que→ Fable 5.1 a diagnostiqué un crash inexpliqué depuis quatre à cinq ans (AFFIRMATION, 0.9)

---
Canonical: https://www.thekb.eu/en/fiches/anthropic-claude-fable-5-1-mythos-5-1-2026-09-01/