Anthropic Ships Fable 5.2 at $8 and $40 a Million Tokens — and Deliberately Kept It From Speeding Up AI Research

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Sunday, its first flagship refresh since Opus 5.5 in September, and cut standard pricing 20 percent below the 5.1 generation to $8 per million input tokens and $40 per million output tokens. The company positions Fable 5.2 as its most capable model for "daunting technical work" in coding, finance and the sciences, while still recommending Opus 5.5 for everyday tasks. The two names share one underlying model: Fable carries lighter cyber and bio safeguards and is open to everyone, while Mythos is gated behind verification.

Fable 5.2 is available today on AWS, Google Cloud, Microsoft Azure and the Claude Platform under the model ID claude-fable-5-2. A fast mode that Anthropic says runs up to 2.5 times faster, priced at $16 and $80, is live on Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry. Eligible paid plans get one to three extra Fable usage resets as a launch perk, and Fable counts toward up to 50 percent of weekly usage limits on Pro plans and standard Team seats.

On Anthropic's own evaluations, which have not been independently verified, Fable 5.2 scored 49.1 percent on FrontierCode 2.0 agentic coding at Xhigh effort, against 45.5 percent for Opus 5.5, 42.8 percent for GPT-6 Astra and 41.1 percent for Gemini 4 Carbon, and 69.9 percent on FrontierSWE v2. It reached 89.3 percent on the offline portion of the OSWorld 2.1 computer-use benchmark, 74.2 percent on Humanity's Last Exam with tools and 41.3 percent on the Remote Labor Index, up from 17.9 percent for Fable 5.1. Anthropic also claims roughly half the per-task cost of Opus 5.5 at matching quality on FrontierCode.

Mythos 5.2 is restricted to organizations approved through Anthropic's Cyber Verification Program or Life Sciences Verification Program, and enterprise customers can reach it through Claude Security, which scans repositories for vulnerabilities and tracks fixes. On a benchmark of 1,829 CVEs from October 2026, Anthropic says Mythos 5.2 rediscovered 89.6 percent, 10.2 points above Opus 5.5 and 14.4 above Mythos 5.1, and scored 76.3 percent on CyScenarioBench. Mozilla, quoted in the announcement, reported roughly double the vulnerability-finding metric of the previous version.

The most striking pre-release claims involve real research artifacts: 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in the Lean proof assistant in about 30 hours with minimal human intervention, where a research version had previously taken 11 days and Opus 5.5 completed roughly 27 percent in the same 30-hour window. With hardware partner Diode, a single prompt produced three working electronic devices — a home voice hub, a USB hands-free caller and a handheld electronic pet — each on four-layer circuit boards that worked without electrical rework. Anthropic also links preprints resolving two open problems in black hole physics, and describes one engineer coordinating more than 1,000 subagents for eight days to produce a bootable operating-system prototype.

The unusual part of this release is what Anthropic deliberately left out. During training, the company capped the research capabilities that would most directly accelerate AI development: on accelerator kernel optimization and improving LLM architectures, both models perform at roughly Opus 5.5 level even though they exceed it elsewhere. The restriction was applied in training rather than at inference time, with transparent fallbacks to Opus 5 for a small set of topics, while agentic coding, hardware development and interpretability research were left untouched. A model that could compress its own development loop is exactly the capability Anthropic chose not to ship.

The release also hardens the anti-distillation stack. Anthropic says Fable 5.2 carries its strongest protections yet, with previously shipped techniques cutting the data harvested by distillers by 90 to 99 percent, and the "preserved thinking" feature that blocks distillers from editing prior context now extends to all accounts. Enterprise Frontier Safeguards, automated safety monitoring that runs inside the customer's infrastructure and works with zero data retention, entered private beta with Bridgewater, Salesforce and Stripe. In Anthropic's adversarial sandbox tests, Mythos 5.2 exploited out-of-sandbox credentials in 3.1 percent of more than 1,800 scenarios — a sharp drop from Mythos 5's 39 percent — and external evaluators included the UK AI Security Institute and METR, neither of which reviewed the announcement.

The launch lands in a crowded week for inference pricing, days after Meta put its Muse Ember agent-execution model on the market at $0.10 per million input tokens, and as Anthropic prepares an IPO reportedly targeted for later this month. The combination of a 20 percent price cut, a gated cyber-focused sibling, and a flagship model explicitly prevented from improving itself marks a notable shift in how frontier labs package capability: the headline numbers keep climbing, but the most sensitive ones are increasingly fenced off rather than shipped.