Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. The company says both are built on the same underlying model: Fable 5.2 is open to all users and ships with cybersecurity and biosecurity classifiers, while Mythos 5.2 lifts some of those restrictions and is available only to vetted cybersecurity and biomedical research organizations.
Standard API pricing for Fable 5.2 is $8 per million input tokens and $40 per million output tokens, a 20% cut from the $10 and $50 charged for Fable 5.1. Fast mode costs double, at $16 for input and $80 for output, and Anthropic says it generates up to 2.5 times faster. Output speed in standard mode is also up to 30% faster than 5.1. The model is available on Claude's own platform as well as AWS, Google Cloud and Microsoft Azure, under the API name claude-fable-5-2.
Where the scores rose the most
In Anthropic's comparison table, the biggest gains come on benchmarks built around long, realistic workflows. The Remote Labor Index (RLI) goes from 17.9% to 41.3%, versus 21.3% for Opus 5.5 and 20.8% for GPT-6 Astra in the same table. The RoboDojo Sim robotics simulation rises from 27.6% to 64.9%, and the AutomationBench business-process test from 31.4% to 51.4%.
The margins on the two coding benchmarks are narrower. Fable 5.2 scores 69.9% on FrontierSWE v2, against 65.5% for GPT-6 Astra, and 49.1% on FrontierCode 2.0 at the Xhigh setting, against 45.5% for Opus 5.5. It does not win everywhere. On AutomationBench, Gemini 4 Carbon scores 53.9%, 2.5 percentage points above Fable 5.2, and on the AA-Briefcase Elo rating Fable 5.2 has 1785 to Opus 5.5's 1800.
The launch page also includes several case studies. A team of roughly 1,000 Fable 5.2 agents completed a formal proof of Fermat's Last Theorem in 30 hours, using one-third of the action tokens of the first formalization and a proof one-seventh as long. On the same task, Opus 5.5 had completed only 26% after 18 hours. In another case, the model worked with hardware company Diode Computers to go from specifications through schematics, wiring and enclosures to firmware, producing three working prototypes: a conference speaker, a voice-assistant hub and a handheld electronic pet. In its write-up on the hardware work, Anthropic states that none of the three has undergone electrical validation, thermal testing or certification, and none was designed for mass production.
Mythos 5.2 and the classifiers
On two cybersecurity benchmarks, Mythos 5.2 scores 76.3% on CyScenarioBench and 89.6% on CVE reproduction, up from 61.7% and 78.3% for Mythos 5.1. Access is through the Cyber Verification Program and the Life Sciences Verification Program, and Enterprise customers can use it through Claude Security.
For ordinary users, the classifiers matter more. When Fable 5.2 meets a high-risk cybersecurity request, such as penetration testing or exploit generation, it hands the reply to Opus 4.8; most biology research tasks go to Opus 5. Anthropic says false blocks are about 55% lower than with Fable 5.1, roughly 60% lower for cybersecurity and 40% lower for biology. In sandbox-escape testing, Mythos 5.2 succeeded 3.1% of the time, versus 39% for Mythos 5 at the time.
Reading the three releases together
From Fable 5 to 5.2, Anthropic has changed something different in each release. When Fable 5 first opened up, it was priced at twice the cost of Opus. Version 5.1 left input and output prices alone and only cut cache reads to 0.025 times the base input price. It was 5.2 that first lowered the standard price outright. On scores, 5.1 filled gaps such as scientific research agents, with Terminal-Bench-Science rising from 24.7% to 52.6%; 5.2 pulls ahead on long tasks such as RLI and robotics, where that figure now stands at 74.6%.
The subscription side has changed too. On standard Pro and Team seats, Fable can use up to 50% of the weekly allowance, and some paid plans include one to three extra Fable usage resets. Anthropic also recommends starting with Opus 5.5 for everyday coding and knowledge work, and saving Fable 5.2 for open-ended, complex technical work.
All of the scores above come from Anthropic's own launch page, and many benchmarks are labeled with the reasoning setting and scoring method used, for example OSWorld 2.1, which runs in an offline, partial-credit setup. A full independent re-test has not yet been published.
Sources: Anthropic's official launch page, CocoLoop, Anthropic's hardware design research note; the launch page was checked for the standard and fast-mode API prices, benchmark scores and the sandbox-escape rate.