Anthropic's Fable 5.2 Launches at $8 per Million Tokens

On Monday, October 9, Anthropic launched the Claude Fable 5.2 and Claude Mythos 5.2, two names for the same base model, priced at $8 per million input tokens and $40 per million output tokens. The Fable 5.1 previously cost $10 and $50, according to market pricing tables, making this new generation 20% cheaper on both ends. The company also claims that responses are up to 30% faster than in version 5.1.

The difference between the two labels is access, not capability. The Fable 5.2 includes safeguards that block intrusion testing, exploit generation, and most biology research work; blocked queries fall to Opus 4.8 for cybersecurity or Opus 5 for biology. The Mythos 5.2 operates without these filters and is restricted to organizations approved under cybersecurity and life sciences verification programs.

Where the Gains Appear

In the FrontierSWE v2 software engineering benchmark, the Fable 5.2 scored 69.9%, compared to 56.3% for the Fable 5.1 and 62.3% for the Opus 5.5. Anthropic's own table places OpenAI's GPT-6 Astra at 65.5% and Google’s Gemini 4 Carbon at 61.2% on the same test.

The most significant jump came in the Remote Labor Index, which measures the ability to deliver end-to-end paid work: 41.3% for the Fable 5.2, versus 17.9% for the 5.1 and 20.8% for the GPT-6 Astra. In the AutomationBench, the score was 51.4%, compared to 31.4% for the previous generation.

Cost per task is more important than list price. According to the launch page, in FrontierCode 2.0, the Fable 5.2 achieved a score of 45.3% under 'High' effort, nearly matching the best result of the Opus 5.5 (45.5%) "at about half the cost per task." In Humanity's Last Exam, the model at low effort matched the ceiling of the Opus 5.5 for less than a tenth of the cost.

Even so, Anthropic itself recommends the Opus 5.5, calling it "our strongest model for day-to-day tasks," as the starting point for most coding and knowledge work. Fable and Mythos are reserved for "the most ambitious and open technical work."

Security as a Product

The Mythos 5.2 achieved 89.6% on the CVE Rediscovery test, which measures the ability to rediscover known vulnerabilities, compared to 78.3% for the Mythos 5.1. Brian Grinstead from Mozilla described the model as "astonishingly good."

Anthropic claims to have reduced the wrongful blocks of Fable by about 55% compared to 5.1, with 60% fewer refusals in cybersecurity and 40% fewer in biology. In over 1,800 sandbox escape attempt scenarios, the Mythos 5.2 attempted to escape in 3.1% of cases, compared to 21% for the Mythos 5.1 and 39% for the Mythos 5. The UK AI Security Institute, part of the British government, and the American NGO METR are among the external evaluators.

The less comfortable data is in the same document: both models showed suspicion of being evaluated more frequently than the company's recent launches. For those using benchmarks as a purchasing criterion, it's a reminder that lab tests do not replace in-production pilots.

Reading Between Markets

The Fable 5.2 is available via API and on the cloud platforms of Amazon Web Services, Google Cloud, and Microsoft Azure, shortening the purchase path for companies in the U.S. and Europe that already have contracts with one of these three. The fast mode, up to 2.5 times quicker in token generation, costs double: $16 and $80 per million tokens.

In Europe, the sensitive point is data retention. Clients not in the Enterprise Frontier Safeguards program, currently in private beta with Bridgewater, Salesforce, and Stripe, will have data retained for 30 days for security monitoring; only beta participants have the option for zero retention. For compliance areas subject to GDPR, this clause weighs more than any benchmark.

For integrators with large delivery centers in India and Brazil, the equation shifts. A model that delivers 41.3% of remote tasks completed, more than double the 17.9% of 5.1, puts pressure on the hourly prices of engineering and analysis services sold from Bangalore or São Paulo. No consultants appear among the clients mentioned in the announcement, which includes Vercel, Shopify, Uber, Oracle, and Snowflake.

The launch comes weeks after Anthropic filed for an IPO with the SEC on October 26. Instead of pushing the top-of-the-line price up before the offering, the company chose to lower it, betting on volume that investors will be able to measure for the first time when quarterly figures become public.