The Day in AI

• Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, one model priced at $8 per million input tokens and $40 per million output tokens.
• Anthropic says the 5.2 models formalized the proof of Fermat's Last Theorem in prerelease testing, reportedly with 1,000 agents working for 30 hours.
• Meta released Muse Ember at $0.10 per million input tokens, and Artificial Analysis scores it 35.8 against 36.2 for Claude Haiku 5.5.
• Google announced TPU 8i Max, two TPU 8i chips in one package, and TPU 8tx, a system for confidential inference.
• Adobe signed a multiyear deal with Disney to build generative models on Adobe Firefly Foundry for Disney's production pipeline.

Anthropic ships Fable 5.2 and Mythos 5.2 as the same underlying model with two access rules. Fable carries cybersecurity and biology safeguards so any customer can use it. Mythos drops those safeguards and goes only to organizations verified through Anthropic's Cyber and Life Sciences Verification Programs. Anthropic reports better coding, reasoning and vision than 5.1, clearer writing, and output up to 30% faster. The context window is 1 million tokens with up to 128,000 tokens of output.

An optional fast mode raises output speed up to 2.5 times, at $16 per million input tokens and $80 per million output tokens. On its Chartography benchmark, Anthropic reports 66.4% for Fable 5.2 at Medium effort. That beats Opus 5.5 at Max effort, which scored 64.4%.

Anthropic also says the models formalized the proof of Fermat's Last Theorem and solved two open black hole physics problems in prerelease testing. A formalization turns a human proof into code that a proof checker verifies line by line. Techmeme's summary of the announcement says the Fermat run used 1,000 Fable 5.2 agents for 30 hours, and that Opus 5.5 with the same setup finished 26% in 18 hours. The announcement page confirms the formalization but does not give the agent count or the duration. On 6 November, Physical Superintelligence said a non-public Anthropic model proved the Aaronson-Ambainis conjecture, and nobody has said publicly whether that model was 5.2.

Meta released Muse Ember, its smallest and cheapest Muse model, aimed at high-volume agent work. It costs $0.10 per million input tokens and $0.50 per million output tokens, has a 1 million token context window, and takes text, images, audio and video. Ember is the fourth Muse model and the first trained using Muse Wonder Full. Artificial Analysis scores it 35.8 on its Intelligence Index, close to Claude Haiku 5.5 at 36.2. It estimates Ember's cost per task at $0.136, about a third of Haiku's $0.411. Haiku leads on FrontierCode 2.0-AA, 34.9% to 27.2%. Ember leads on AutomationBench-AA, 37.8% to 13.9%.

• Google says TPU 8i Max doubles compute, memory and bandwidth over TPU 8i, though the specs could only partly be checked. It also claims TPU 8tx has up to 2x lower latency than TPU 8t.
• Adobe and Disney will build models trained on Disney characters, environments and styles for theme parks, the cruise line, Walt Disney Studios and ESPN.
• Nous Research released Arete-27B, a fine-tune of Qwen3.8-27B that it says cut party-line answers from 22% to 1% on its own 803-prompt eval.
• DeepSeek reportedly has 1.54 million daily users on DeepSeek Harness, its open-source agent framework released in August.
• Sashikumar Sreedharan, a senior Google Cloud executive, reportedly joined Sarvam AI as President, Chief Business Officer and board member.
• Anthropic settings files from Claude Code are exposing thousands of secrets on GitHub, according to Ars Technica.
• Fireworks AI suffered a new hack of customer data, according to The Information.
• Nvidia intervened before a Commerce Department chip-smuggling investigation into Megaspeed moved to the Justice Department, according to The New York Times.
• The European Commission has drafted a Digital Fairness Act that would target AI systems simulating relationships, according to Bloomberg.