Claude Fable 5.2: Faster, Cheaper Frontier Coding for Business

Claude Fable 5.2 is Anthropic's new frontier coding model, launched on 9 November 2026. Pricing is $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1 — a 20% cut — while output is up to 30% faster. For a business buying coding agents, that matters more than any benchmark.

Fable 5.2 is the same underlying model as Mythos 5.2, with one difference: Fable's safeguards block most cybersecurity and biology work, so Anthropic can offer it to everyone, while Mythos drops those blocks for vetted organisations only. It is available on Pro, Max, Ultra, Team and Enterprise plans, on the Claude Platform and in Claude Code, and through AWS, Google Cloud and Microsoft Foundry. Anthropic's guidance: start with Claude Opus 5.5 for most coding and knowledge work; Fable is for the most ambitious, open-ended technical work.

What Claude Fable 5.2 actually is

Fable 5.2 is built for long, open-ended technical work: multi-day coding projects, unsupervised agent runs, and research across coding, science and computer use — tasks too big for a standard coding assistant, where the model must plan, try, fail, recover and keep going. For ordinary day-to-day code, Opus 5.5 is the recommended starting point, and the cheaper call. Fable 5.2 earns its place on the hardest jobs: migrating a legacy service, building a system from a vague brief, or running a fleet of agents on one problem. Our guide to the chatbot-versus-agent choice makes the same distinction, and our AI agents and automation work starts from that question.

Fable can also act as an Advisor, where smaller models consult it, or an Orchestrator that delegates to smaller models — keeping the expensive model where its judgement matters. That is the same cost discipline we apply when we scope AI agent development costs for clients.

What it costs

The headline price is $8 per million input tokens and $40 per million output tokens, down 20% from Fable 5.1, with output up to 30% faster. An optional fast mode at $16/$80 per million tokens offers up to 2.5x token generation speed. Standard mode is the value option; fast mode only pays when speed is the bottleneck — a live demo, a deadline, or an agent loop where waiting costs more.

Cheaper frontier models are a theme this month: our coverage of Meta's Muse Ember shows the same pressure, and LG's K-EXAONE 2.5 free API is pushing costs down further still. The businesses that benefit most already know which jobs to automate — our custom software cost guide for the UK shows how to budget the build.

The benchmark scores — read them as Anthropic's claims

Anthropic's benchmark table puts Fable 5.2 ahead of its predecessor and rivals. FrontierCode 2.0 is 49.1%, against 32.4% for Fable 5.1, 45.5% for Opus 5.5, 42.8% for GPT-6 Astra and 41.1% for Gemini 4 Carbon. FrontierSWE v2 is 69.9%, against 56.3% for 5.1 and 65.5% for GPT-6 Astra. Not every score leads: Gemini 4 Carbon tops AutomationBench at 53.9% against Fable 5.2's 51.4%. Terminal-Bench-Science is 74.6%, OSWorld 2.1 computer use is 89.3%, RoboDojo Sim robotics is 64.9% — up from 27.6% — and Humanity's Last Exam with tools is 74.2%.

These are Anthropic's claims, not independently verified. First, do not confuse these scores with the GDPx benchmark we covered on 2 November, where the best agent cleared only about 12% of long-horizon tasks; that is a different test, which Fable's scores do not mean it has solved. Second, the only score that counts is the one you run yourself, on your own code.

The headline demos: what Anthropic says 1,000 agents did

Anthropic says 1,000 Fable 5.2 agents formalised Fermat's Last Theorem in 30 hours — a proof whose first formalisation took 11 days — and solved two open problems in black-hole physics. These are Anthropic's claims, not independent findings. The hardware demos are the most concrete: Anthropic says Fable 5.2 designed three working devices with Diode Computers — a speakerphone, a voice-assistant hub and a handheld virtual pet — boards ready in under a day. One engineer plus 1,000 subagents built a booting prototype operating system in eight days, and a team migrated a critical production service from Python to Rust, now starting 26 times faster.

Treat the demos as a direction, not a promise. A thousand-agent run is a research showcase, and coordinating it is itself the hard part — which is why our coverage of Cognition's SWE-2 Flash matters. But the Python-to-Rust migration is the demo a business owner should study: the closest to a real job.

Where it fits in a real business

Fable 5.2 fits work that is too big, too long or too uncertain for a normal coding assistant:

• Legacy modernisation: migrating services between languages or frameworks, and paying down technical debt no sprint gets to.
• Greenfield builds from a vague brief: turning a rough product idea into a working system. Our custom software development work starts from briefs like this.
• Agent fleets: running many agents in parallel, with Fable as Orchestrator — see our LongCat-2.5 coverage.
• Research and science: coding-heavy research, data analysis and simulation — see our AGIBOT GO-3 coverage and Nvidia's GR00T 2.

It does not fit everywhere. Short, well-specified tasks are cheaper on Opus 5.5 or smaller models, and cybersecurity or biology work will hit Fable's safeguards by design. Businesses handling regulated data should read Anthropic's usage rules first — our coverage of Anthropic's November usage-policy change explains what changed.

Safety, Mythos and the IPO in the background

Anthropic says Fable 5.2's safeguards are as robust as 5.1's but with fewer unnecessary blocks, with the strongest protections yet against distillation attacks. For a business, the safeguards are a feature, not a bug.

Mythos 5.2 is the cyber story: Anthropic reports CyScenarioBench at 76.3%, against 61.7% for Mythos 5.1 and 61% for GPT-6 Cyber, CVE rediscovery at 89.6%, and 10.2% more vulnerabilities found than Opus 5.5 — Anthropic's own figures; Mythos is for vetted organisations only. Around the launch, Reuters and The Hindu reported Anthropic is expected to list in an IPO in the week of 16 November, which could raise up to $100 billion — reported as the largest IPO on record if it happens. Treat that as reported, not confirmed. Read the Anthropic announcement and the Reuters report.

What your business should do now

Do not rebuild your engineering process around a model that is one day old. Run a contained test instead: pick one real job from your backlog — a migration, a dependency upgrade, a service you have been meaning to rewrite — and run Fable 5.2 on it in a branch, with your own tests as the judge. Score it like a new hire: did it pass your tests and save review time?

Then compare the cost honestly. At $8 and $40 per million tokens, a long agent run is cheaper than on 5.1, but a runaway agent can still burn through tokens fast — set budgets and step limits first, and use the Advisor and Orchestrator patterns to keep the flagship model off routine work. Our AI voice calling agent guide applies the same pilot discipline: one narrow pilot, measured properly, beats a company-wide rollout on launch-day hype. If you want help scoping a pilot, talk to Coheser about an AI agent pilot or contact us.

Frequently asked questions

What is Claude Fable 5.2?

Anthropic's frontier coding model, launched 9 November 2026 — the same underlying model as Mythos 5.2, with safeguards blocking most cybersecurity and biology work, aimed at the most ambitious, open-ended technical work.

How much does Claude Fable 5.2 cost?

$8 per million input tokens and $40 per million output tokens, down 20% from $10 and $50 for Fable 5.1. An optional fast mode at $16/$80 offers up to 2.5x token generation speed.

Should my business use Fable 5.2 or Opus 5.5?

Start with Opus 5.5 for most coding and knowledge work, says Anthropic; use Fable 5.2 for the most ambitious, open-ended technical work.

Are the benchmark scores independently verified?

No. All figures here are Anthropic's own claims. Test the model on your own code and acceptance tests first.

What is the difference between Fable 5.2 and Mythos 5.2?

They are the same model. Fable's safeguards block most cybersecurity and biology work; Mythos lacks those blocks and is for vetted organisations only.