Claude Fable 5.2: Anthropic's new model tops Opus 5.5 at coding and costs 20% less

Claude Fable 5.2 is Anthropic's new top public model, released on November 9, 2026, alongside a restricted twin called Claude Mythos 5.2. Anthropic says Fable 5.2 is the best released model it has tested on agentic coding, answers up to 30% faster than Fable 5.1, and costs 20% less: $8 per million input tokens and $40 per million output tokens, according to Anthropic's announcement and CNBC.

Our 27-second version:

New to AI agents? Read: What are AI agents?

The diff in one line: until this launch, Anthropic's public Fable 5.1 trailed Opus 5.5 on Anthropic's hardest coding test. Now Fable 5.2 leads it, at a lower price than its predecessor.

What is Claude Fable 5.2?

Fable and Mythos are the same underlying model. The difference is the guardrails. Fable ships with strong cybersecurity and biology safeguards, which is why Anthropic can offer it to everyone. When a request trips those safeguards, Fable quietly hands it to an older model: Opus 4.8 for cybersecurity and Opus 5 for biology.

Mythos 5.2 has fewer of those blocks and is meant for vetted organisations. Security teams apply through Anthropic's Cyber Verification Program and life-science researchers through its Life Sciences Verification Program. Enterprise customers also get Mythos 5.2 inside Claude Security.

Anthropic says Fable 5.2 hits its safeguards about 55% less often than Fable 5.1 did, with roughly 60% fewer blocks on cybersecurity work and 40% fewer on biology, per MacRumors.

Image: Anthropic (press video still)

Claude Fable 5.2 benchmarks: how does it compare with Opus 5.5?

On FrontierCode 2.0, Anthropic's agentic coding test, Fable 5.2 scored 49.1% at its "Xhigh" effort setting. Fable 5.1 scored 32.4%, Opus 5.5 45.5%, OpenAI's GPT-6 Astra 42.8% and Google's Gemini 4 Carbon 41.1%. At "High" effort Fable 5.2 lands level with Opus 5.5 at roughly half the cost per task, Anthropic says.

Chart: NextDiff, from scores Anthropic published (vendor-reported)

The biggest jumps are in long, real-world tasks:

• Remote Labor Index (paid freelance jobs): 41.3%, up from 17.9% for Fable 5.1 and about double Opus 5.5 (21.3%).
• RoboDojo Sim (controlling robots): 64.9%, against 39.4% for Opus 5.5.
• OSWorld 2.1 (using a computer): 89.3%, ahead of Opus 5.5 at 87.2%.

It does not win everything. Gemini 4 Carbon scores higher on AutomationBench (53.9% vs 51.4%), and Opus 5.5 keeps a slightly higher Elo on the AA-Briefcase knowledge-work test. Independent tester Artificial Analysis puts Fable 5.2 first on its Intelligence Index with a score of 65, but also calls it slow, verbose and the most expensive model per task it measures ($26.54 per index task).

What has Fable 5.2 actually built?

Anthropic's launch post leans on demos rather than charts:

• An operating system: one engineer had Fable 5.2 coordinate more than 1,000 sub-agents for eight days, and the result boots on a laptop.
• A maths proof: a team of 1,000 Fable 5.2 agents formalised Fermat's Last Theorem in the Lean proof language in about 30 hours.
• Hardware: with the startup Diode, it designed three electronic devices, including a voice-assistant hub, that worked without electrical fixes.
• Robots: it controlled a pair of physical robot arms.

Image: Anthropic (press video still)

These are Anthropic's own showcase projects, picked by the company. They show what is possible with large budgets and many agents, not what a single chat will do.

How much does Claude Fable 5.2 cost, and who can use it?

• API price: $8 per million input tokens and $40 per million output tokens (model id claude-fable-5-2). CNBC reports that is 20% below the 5.1 generation.
• Fast mode: up to 2.5 times faster output for $16 / $80 per million tokens, on Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry.
• Where: Anthropic's apps and API, plus Amazon Web Services, Google Cloud and Microsoft Azure.
• Subscribers: Fable can use up to 50% of the weekly usage limit on Pro plans and standard Team seats, and paying subscribers get one to three Fable "resets" depending on how long they have subscribed.

Anthropic still recommends Opus 5.5 as the everyday model for most coding and writing. Fable is pitched at "power users or power use cases", as product lead Dianne Penn told CNBC.

Why it matters

Price per task is now the real benchmark. Fable 5.2 is not cheap per token, but if it finishes a job in fewer steps than Opus 5.5, the bill can still fall. For teams running coding agents, that is the number to test on their own work.

The timing matters too. CNBC reports that Anthropic is preparing an IPO that could come as soon as next week, at a valuation of $1 trillion to $1.5 trillion. We covered that shift in Anthropic's IPO valuation on Sunday. A flagship launch days before listing is also a pitch to investors.

What's next and the caveats

Every score above comes from Anthropic and has not been independently reproduced, apart from Artificial Analysis's own index. The system card has more safety detail; Anthropic rates the risk of the models acting unsafely in high-stakes situations as low. Expect head-to-head tests against GPT-6 Astra and Gemini 4 Carbon in the coming days. New to building apps with AI? Read our explainer on what vibe coding is.