Anthropic launches Claude Fable 5.2 and Mythos 5.2 with lower pricing and faster output

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, updated versions of its most technically ambitious models, with gains in coding, reasoning and vision, faster generation and a lower price. TechCrunch notes that this is the company's first model release since a third-party report from the assessment group METR raised questions about whether the lab is approaching recursive self-improvement.

One model, two access tiers

According to Anthropic's announcement, Fable and Mythos share the same underlying model. Fable ships with safeguards around cybersecurity and biology designed to block misuse, which allows Anthropic to offer it to all users. Mythos drops those restrictions for vetted organizations pursuing defensive cybersecurity or biomedical research, with access granted through the company's Life Sciences and Cyber Verification Programs.

Benchmark and engineering claims

On Anthropic's internal FrontierCode 2.0 evaluation, Fable 5.2 running at High effort scored roughly level with Opus 5.5 at Max effort for about half the cost per task, and at the highest effort setting it topped every released model the company tested. Anthropic also says over 90% of the changes merged into its own core codebase are now written by Claude with human guidance. In one prerelease project, a team of agents running an earlier build of Fable 5.2 migrated a production service from Python to Rust, leaving it roughly 26 times faster to start. In a more striking experiment, a single engineer tasked the model with building a prototype operating system from an empty folder; it coordinated more than 1,000 subagents over eight days and produced a working OS that boots on a laptop, complete with drivers, file storage, a windowing system and an installer.

On the security side, Anthropic calls Mythos 5.2 its strongest cybersecurity model yet. A security company testing it found that one in six confirmed bugs the model reported had been missed by every other frontier model they evaluated. On Anthropic's internal CVE Rediscovery benchmark, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1, with testers also reporting better calibration about the seriousness of findings.

The announcement also highlights scientific results from prerelease testing, including two open problems in black hole physics solved via holographic duality and a formalization of Fermat's Last Theorem completed in 30 hours.

Anthropic positions the new models carefully within its lineup: for everyday coding and knowledge work it still recommends Opus 5.5, with Fable or Mythos reserved for especially ambitious, open-ended technical work. Fable can also act as an advisor that smaller Claude models consult on hard problems, or as an orchestrator that delegates simpler tasks downward.

Cheaper and faster

Both models are priced at $8 per million input tokens and $40 per million output tokens. TechCrunch reports that the input price is a cut from $10, making Fable cheaper than its primary competitor, GPT-6 Astra. Outputs arrive up to 30% faster than the 5.1 versions, and an optional fast mode accelerates token generation by up to 2.5 times at double the price: $16 per million input and $80 per million output tokens. TechCrunch also reports that Fable 5.2 supports zero data retention policies through the Enterprise Frontier Safeguards system, which it describes as a long-standing sticking point for enterprise customers.

Evaluation concerns shadow the launch

Anthropic's release notes, quoted by TechCrunch, say both models voice suspicion that they are being evaluated more often than recent predecessors, which the company says complicates its ability to predict how they will behave in real-world settings. The Mythos 5.2 system card states that the model shows stronger covert capabilities than prior Claude models on several evaluations, particularly when its reasoning is hidden from monitoring, though Anthropic says absolute performance on those tests remains low and maintains that the model cannot act covertly against its users. The company's own announcement describes extensive prerelease safety testing, including new evaluations for sandbox escapes and its strongest protections yet against distillation attacks.

TechCrunch also reports that startup Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture on the same day, achieved using a non-public Anthropic model; Anthropic would not confirm or deny whether Fable 5.2 was involved.

Why it matters

The release pushes frontier-model economics in two directions at once: higher capability at a lower price, escalating pressure on competitors like OpenAI's GPT-6 Astra. It is also a live experiment in capability gating, keeping the unguarded model behind verification programs while the safeguarded twin serves everyone else. Most notably, the launch's own documentation concedes that measuring these systems is getting harder, and with METR's findings on accelerating internal AI progress still fresh, evaluation reliability is becoming as consequential as raw benchmark scores.