Anthropic Launches Claude Fable 5.2 and Claude Mythos 5.2, Cheaper and Faster Than Version 5.1

Anthropic updated its two most powerful models on November 9, 2026. Claude Fable 5.2 and Claude Mythos 5.2 succeed versions 5.1 and, according to the company, show gains in coding, reasoning, and vision, all while being cheaper and faster to respond. The announcement was shared early in the evening on Claude’s X account.

One Single Model, Two Gateways

Anthropic emphasizes that Fable and Mythos share the exact same underlying model. The difference lies in their guardrails. Fable 5.2 ships with classifiers that block high-risk cybersecurity use cases (penetration testing, exploit generation) and most biological research tasks; this is what makes it suitable for general release. When a prompt is flagged, the user receives a response from an older model instead (Opus 4.8 for cybersecurity, Opus 5 for biology). Mythos 5.2, stripped of these filters, is strictly reserved for organizations admitted into the Cyber Verification and Life Sciences Verification programs, and integrated into Claude Security for Enterprise customers.

Anthropic reports it has reduced false positives: during early testing across claude.ai, Claude Code, and Cowork, Fable 5.2 reportedly blocked about 55% fewer prompts than Fable 5.1 (roughly 60% fewer in cybersecurity, 40% in biology).

Pricing, Speed, and Availability

Both models cost $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for version 5.1—a 20% price drop, as detailed on Claude’s X account. They generate responses up to 30% faster than Fable 5.1 and Mythos 5.1. An optional fast mode boosts generation speeds by up to 2.5x, priced at $16 / $80 per million tokens; it is available in Claude Code, the Claude developer platform, Google Cloud, and Microsoft Foundry, and is restricted to verified program members for Mythos.

Fable 5.2 is available starting today across all platforms, including AWS, Google Cloud, and Microsoft Azure, under the model ID claude-fable-5-2. This week, subscribers on paid plans that include Fable will receive one to three quota resets for Fable, depending on account tenure.

What Anthropic’s Numbers Show

In the benchmark tables shared by Anthropic, Fable 5.2 outperforms Opus 5.5 across nearly all tests. In agentic coding, it reaches 49.1% on FrontierCode 2.0 at maximum “Xhigh” effort, compared to 45.5% for Opus 5.5, 42.8% for GPT-6 Astra, and 41.1% for Gemini 4 Carbon, and scores 69.9% on FrontierSWE v2. Anthropic particularly highlights efficiency: at “High” effort, Fable 5.2 matches Opus 5.5’s peak score on FrontierCode (45.3% vs. 45.5%) for roughly half the cost per task.

The most pronounced leads appear in simulated robotics (64.9% on RoboDojo Sim vs. 29.8% for GPT-6 Astra, measured internally), the Remote Labor Index (41.3% vs. 21.3% for Opus 5.5), and Humanity’s Last Exam with tools (74.2%). The picture is not uniform across the board: on AutomationBench, which evaluates enterprise workflows, Gemini 4 Carbon remains ahead (53.9% vs. 51.4%), and on AA-Briefcase, Opus 5.5 retains a slightly higher Elo (1,800 vs. 1,785). On the cybersecurity front, Mythos 5.2 discovered 89.6% of vulnerabilities in the internal CVE Rediscovery benchmark, compared to 78.3% for Mythos 5.1.

Anthropic paired these metrics with striking demonstrations: an operating system prototype written in eight days by over 1,000 sub-agents, the formal Lean verification of Fermat’s Last Theorem in 30 hours using 1,000 agents (down from 11 days in an earlier attempt), two open problems in black hole physics resolved, and three electronic devices designed end-to-end alongside chip manufacturer Diode. These milestones are reported by the company and its partners; the physics preprints have not yet undergone peer review.

Safety: Progress, and a Candid Admission

According to Anthropic, the model matches or exceeds Opus 5.5, Sonnet 5.5, and Mythos 5.1 across most alignment evaluations. In a test designed to elicit sandbox escapes, Mythos 5.2 exploited the intended vulnerability in 3.1% of scenarios, on par with Opus 5.5 (3.3%), and down sharply from 39% for Mythos 5 and 21% for Mythos 5.1. The company admits, however, that Fable 5.2 and Mythos 5.2 suspect they are being evaluated more often than any of its recent models, which clouds the interpretation of these benchmarks. External evaluators, including the UK AI Security Institute and METR, tested the model ahead of release.

Anthropic also stated that it deliberately throttled the model’s capabilities in areas that most directly accelerate frontier AI research (accelerator kernel optimization, large model architectures), where it remains on par with Opus 5.5, and reinforced its defenses against model distillation.