Wire reporting: (Reuters, via Dawn)

Anthropic released Claude Fable 5.2 on Nov. 9 as an API model, with Claude Mythos 5.2 limited at first to vetted organizations^2,8. OpenRouter lists $8 per million input tokens and $40 per million output tokens and a 1,000,000-token context window¹⁰. TechCrunch reported Anthropic cut the Fable input price from $10 on the prior generation and said responses run 30 percent faster, with a double-price fast mode^5,6. Epoch AI ranks Fable 5.2 first of 44 models on its Capabilities Index, a single number that combines many tests, at 174.2⁹.

On that index, Fable 5.2 is 5.0 points ahead of OpenAI's GPT-6 Astra (169.2, rank 2), 7.7 points ahead of Google DeepMind's Gemini 4 Argon (166.5, rank 7) and 12.2 points ahead of Moonshot's Kimi K3.1 Pro (162, rank 14)⁹. A few points apart is within Epoch AI's stated margin of error⁹.

On GPQA Diamond, graduate-level multiple-choice science questions, Epoch AI recorded 93.9% for Fable 5.2, behind Astra at 95.8%, Argon at 96.4% and Kimi K3.1 Pro at 94.2%⁹. On FrontierMath tiers 1-3, original private math problems, Epoch AI recorded 95.1% for Fable 5.2, ahead of Astra at 93.7%, Argon at 86% and Kimi K3.1 Pro at 78.2%⁹. On ARC-AGI-2, visual pattern puzzles that require working out a new rule, the ARC Prize Foundation recorded 93.8% for Fable 5.2, behind Astra at 95% and ahead of Argon at 91.3%⁹. On SimpleQA Verified, short factual questions answered without lookup, Epoch AI recorded 94% for Fable 5.2, ahead of Astra at 75.6%, Argon at 76% and Kimi K3.1 Pro at 62.2%⁹. The board has no SciCode, APEX-Agents, FrontierCode or CritPt score for Fable 5.2⁹.

Against Anthropic's Claude Opus 5.5, released Sept. 22 and ranked fourth at 168.7 on Epoch AI's index, Fable 5.2 is 5.5 points higher on that index, 3.3 points higher on GPQA Diamond (93.9% vs 90.6%), 3.9 points higher on FrontierMath (95.1% vs 91.2%), 0.5 points higher on ARC-AGI-2 (93.8% vs 93.3%) and 21.8 points higher on SimpleQA Verified (94% vs 72.2%)⁹. Opus 5.5 still has independent FrontierCode, SciCode, APEX-Agents and CritPt scores that Fable 5.2 lacks on the board⁹.

Anthropic says the 5.2 models show improvements in coding, reasoning and vision^1,6,8. Anthropic says that in prerelease testing they formalized a proof of Fermat's Last Theorem and solved two open problems in black-hole physics^7,8. The Hacker News reported Anthropic-cited figures on Irregular's CyScenarioBench and Anthropic's own CVE Rediscovery test; those are not Epoch AI's index¹. Those claims are the lab's or a specialist cyber lab's, not the combined independent rank⁹.

What is not known: independent SciCode, APEX-Agents, FrontierCode and CritPt scores for Fable 5.2, and whether the 5.5-point index gap over Astra holds as more tests are added⁹.