Anthropic releases Claude Fable 5.2 and Mythos 5.2, its top AI models, 20% cheaper

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2, new versions of its two flagship models. It says they are better at coding, reasoning and vision, and that they write answers up to 30% faster than the 5.1 versions. The New Stack reports that both cost $8 per million input tokens and $40 per million output tokens, 20% less than 5.1. A faster mode costs $16 and $80. Anthropic also says that in testing before release the models wrote a formal, machine-checked version of the proof of Fermat's Last Theorem and solved two open physics problems about black holes. GitHub says Fable 5.2 is now available in its Copilot coding assistant for paid plans. TechCrunch notes the release comes amid concerns about AI that can speed up its own development.

Meta releases Muse Ember, its smallest and cheapest AI model so far

Meta Superintelligence Lab released Muse Ember, its lightest and most cost-efficient AI model, TechCrunch reports. It costs 10 cents per million input tokens. Meta says it gives more intelligence per dollar than small rival models Haiku 5.5 and GPT-6 Luna. TechCrunch describes it as a lightweight model for carrying out agent tasks. Some users on Reddit point out that Meta has again released a closed model rather than open weights.

Nvidia releases GR00T 2, an open AI model that teaches robots delicate hand tasks

At the Conference on Robot Learning in Austin, Nvidia introduced GR00T 2, an open model that helps robots predict how their actions will change their surroundings. Nvidia says it completed 40 of 50 tested tasks at least once without task-specific demonstrations. It says the model reached 71% average success with one demonstration per task and 95% with five. Nvidia also says its Isaac Lab 3.0 simulator with the Newton physics engine trains up to 2.5x faster than Isaac Lab 2.3. Foxconn is applying GR00T 2 to server manufacturing, and Siemens is evaluating it.

Nvidia expands AI work with Korea's government and Korean chip startups

Nvidia says it is expanding its work with Korea's Ministry of Science and ICT, covering AI chips, data centres, training, startups, models and security. Korean chip companies may explore joining NVLink Fusion, Nvidia's system for linking other companies' chips with its own. Nvidia also announced a collaboration with SK Telecom and chip startup Rebellions on AI infrastructure for Korean models. Separately, it said it has invested in Korean chip-link startup Point2 and plans investments in Motif Technologies, RLWRLD and Upstage.

AI company Anthropic pays $3.4B for gas generators to power its data centres

Anthropic has committed $3.4 billion to secure 1.7GW of gas-fired generators from OnePWR Solutions, Tom's Hardware reports. It takes over existing Caterpillar orders, avoiding long waits for new equipment. TechCrunch says the deal could leave Anthropic with buyer's remorse.

Nvidia shares fall as its board and the FBI look into CEO Jensen Huang's affair

Nvidia's shares dropped about 5% in pre-market trading after its board opened an inquiry into CEO Jensen Huang, the Financial Times reports. Huang admitted an affair with Alice Huang, a former executive at Nvidia customer Megaspeed, whom US intelligence suspects of being a Chinese agent. President Trump said the FBI is "looking into" the affair, and that Huang told him he "had absolutely no idea that the person in question was a spy", the Associated Press reports. Huang has been summoned before Congress's Select Committee on the CCP. The New York Times reports, citing sources, that a Commerce Department probe into Megaspeed for allegedly smuggling Nvidia chips to China was moved to the Justice Department after Nvidia intervened. TechCrunch notes a short seller had flagged the relationship in May. Nvidia says it has no reason to believe company secrets were compromised.

Google shows two new AI chips, TPU 8tx and TPU 8i Max, launching in 2027

Google has added two systems to its eighth generation of TPUs, its own AI processors. It says the TPU 8tx keeps data and models protected inside the hardware for sensitive work, and is paired with AMD's 5th Generation EPYC processors. The TPU 8i Max is built for very long inputs and the highest computing density per rack. Bloomberg reports both are due in 2027. Google says this makes four new TPU systems in one year, and that TPUs now have nearly 100 customers.

AI cloud provider Fireworks says hackers reached some customers' data and billing details

Fireworks, which runs AI models for other companies, says hackers accessed some customers' datasets, incomplete model weights and billing information over the past month, Forbes reports. Two customers told Forbes they had already moved most of their business off Fireworks' systems.

France blames Iran's Revolutionary Guards for hacks that knocked out shop fridges

French officials on Monday accused Iran's Revolutionary Guards of being behind cyberattacks that knocked out commercial refrigerators and freezers across France in October, Reuters reports. TechCrunch says France's foreign minister blamed Iran for at least two waves of attacks on refrigerated and freezer units in France and wider Europe, including in grocery stores and blood banks.

Apple starts selling $49.99 packs of extra AI use in its Logic Pro music app

Apple has added a "Buy More Usage" option to Logic Pro for a one-time price of $49.99, 9to5Mac spotted. It applies once users hit monthly caps on some Apple Intelligence features, starting in Apple's Creator Studio apps. Logic Pro has a built-in AI assistant that can do things like start a new song. Apple published usage limits for some AI features in early September and had said it would eventually charge for more use.

Everything else