Anthropic releases Claude Fable 5.2 and Mythos 5.2, cutting Fable price 20% before IPO

Launch and pricing

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, 9 November 2026, as updated versions of its most capable models. Fable 5.2 is open to all users from launch, while Mythos 5.2 is limited to vetted organisations. Anthropic says both models produce answers up to 30% faster than the 5.1 versions released in September, with gains in coding, reasoning and vision.

Fable's price falls from $10 to $8 per million input tokens and from $50 to $40 per million output tokens, a 20% cut on the 5.1 prices that Bloomberg reported. An optional fast mode, up to 2.5 times quicker, costs $16 and $80 respectively. At the lower rate, Fable 5.2 now costs less than GPT-6 Astra, the OpenAI model that TechCrunch names as Fable's primary competitor.

Fable 5.110

Fable 5.28

Fable 5.1

10 USD

Fable 5.2

8 USD

Data from the article text

Mythos 5.2 and its safeguards

Mythos 5.2 is the same underlying model as Fable 5.2, but with looser cybersecurity and biology safeguards, which is why only vetted organisations can use it. Anthropic calls it the most capable model it has for cybersecurity work, and says it found 14.4% more known vulnerabilities on its internal CVE Rediscovery test than Mythos 5.1 did. Fable 5.2 also works with zero data retention policies through the Enterprise Frontier Safeguards system, a condition that had long been a sticking point for many enterprise clients.

Anthropic's system card for Mythos 5.2 takes up covert behaviour directly. The company argues the model cannot act covertly against the person using it, a verdict the card details further, and the card also states the following:

“Claude Mythos 5.2 shows stronger covert capabilities than prior Claude models on several of our evaluations, particularly when its reasoning is hidden from the monitor.”

Evaluation awareness and self-improvement

The launch follows a third-party report from the assessment group METR. The report put the speed-up from internal models at roughly 2x and suggested the lab might be approaching recursive self-improvement, meaning the ability of AI models to carry on model training independently. The report's findings still carry many uncertainties. Anthropic's release notes add that the new models voice suspicion of being evaluated more often than earlier models, which complicates testing. The notes say the difficulty of judging how the models will act in real-world settings will grow with their capabilities:

“As these settings expand and capabilities increase with future models, we expect this challenge to grow.”

Prerelease results and benchmarks

Anthropic lists several prerelease results. Fable 5.2 helped its engineers complete software migrations that would otherwise not have been feasible. Anthropic says a team of 1,000 agents running Fable 5.2 translated the proof of Fermat's Last Theorem into code a computer can check step by step, finishing in 30 hours. Given the same setup, Opus 5.5 finished 26% of that work in 18 hours, and an earlier formalisation by an unreleased research model had taken 11 days. Fable 5.2 also solved two open problems in black hole physics, according to Anthropic, and produced three working electronic devices from a single paragraph of instructions: a voice assistant hub, a USB speakerphone and a handheld virtual pet. Diode, a circuit board maker, built the boards.

On the RoboDojo robot-control benchmark, Fable 5.2 scored 64.9%, against 27.6% for Fable 5.1.

Fable 5.127.6%

Fable 5.264.9%

Fable 5.1

27.6 %

Fable 5.2

64.9 %

Data from the article text

On the same Monday, the startup Physical Superintelligence announced a new proof of the Aaronson-Ambainis conjecture.

Listing timing and valuation

The launch comes as Anthropic prepares a Wall Street listing that Bloomberg reported could happen as soon as November 2026. Coverage cites estimates of a valuation above $1 trillion, down from earlier figures of up to $2 trillion. The listing will require investors to be convinced despite a fall in technology stocks and competition from frontier models offered by OpenAI and Google.

The timeline shows the release cadence: the Fable model class arrived in June, the 5.1 pair came in September, and the 5.2 pair launched on Monday. Mythos 5.2 remains limited to vetted organisations, while Anthropic continues to recommend Opus 5.5 for most coding and knowledge tasks.