AI News Daily

Mistral Large 4 joins the Open Math OS model lineup

Mistral AI announced that Mistral Large 4 has joined the Open Math OS model lineup, featuring 1.05T open weights and a 1M context window, and opening to approved teams starting today. Open Math OS is a research workspace where mathematicians and AI agents collaborate, built solely on open models, so the community can inspect the runs and control its own data.

Apple opens Siri AI to third-party apps, launches App Skills framework

Apple announced that it is opening Siri AI capabilities to third-party apps and launching a new framework called App Skills, which lets developers combine multi-step operations into a single natural-language command for execution. Siri can also complete tasks by combining the current interface context with cross-app coordination, for example finding a product in a shopping app and then sending it via a messaging app. The related capabilities are built on the App Intents framework, require no special permission application or enrollment in a new developer program, and are currently available for testing through the iOS 27.3 developer preview.

FastVideo supports the Open-H3-Regenerate-2K model

Hao AI Lab announced that FastVideo now supports the Open-H3-Regenerate-2K model from HaoZhe65347 and NVIDIA AI, and provided a link to the 2K model. The series previously covered FastVideo's inference optimizations for Pika Labs' open-source MiniMax H3 Flex, which takes 5.8 seconds to generate a 5-second 480p video with audio on a single GB200, up to 7 times faster than the official open-source inference code, with newly added multi-GPU support.

Google shuts down Search for schools that refuse AI features

Google told The New York Times that it will completely shut down Search for schools that refuse its AI features. In an education-related update on Tuesday, Google said that unless administrators explicitly enable it, K-12 student accounts' access to Gemini, Gemini Notebook, Gemini in Classroom, and Google Search "will be completely turned off." Chris Phillips, Google's vice president of education, said Search is deeply integrated with AI and that schools cannot disable Gemini features separately without turning off Search. Google had previously enabled Gemini by default in Google Classroom, which is used by more than 150 million teachers and students worldwide, and last month lawmakers pressured Google over the potential harms of AI products entering schools. Google will also embed the Guided Learning feature into Search's AI Overviews, testing it in some U.S. schools starting this month.

Nvidia requires cloud and data center customers to direct at least 51% of services in new contracts to non-Chinese companies

According to FT, Nvidia has told cloud and data center customers that at least 51% of services in newly signed contracts should go to non-Chinese companies. The requirement targets overseas cloud rental, a path by which Chinese AI developers can use advanced computing power without importing chips, because U.S. export controls generally do not cover remote access. Cloud providers are also required to report the identity of end users of the chips; earlier this year Nvidia drew up a whitelist that cut the number of Asian customers allowed to buy its AI chips to less than half, and a B300 rack sells internationally for about $600,000, with a transaction price in China last month reaching $2.54 million.

WSJ: Large investors want Anthropic's IPO priced at $1.5 trillion to $1.6 trillion

The WSJ reported that large investors want Anthropic's IPO priced at $1.5 trillion to $1.6 trillion, below the nearly $2 trillion discussed in previous months. The revised prospectus shows that Anthropic's third-quarter revenue rose from $1.16 billion to $1.7 billion, but spending grew faster, turning $974 million in operating profit into an operating loss, mainly because $5 billion in computing power procured under short-term terms cost more. Its total long-term commitments amount to about $895.4 billion, almost all for computing power and infrastructure, of which $447.5 billion is expected to be spent before 2031.

Anthropic releases FlashProve, which can verify the source of model replies, and open-sources the code

Anthropic announced FlashProve, a new research effort in the direction of verifiable inference, based on existing cryptographic techniques that let developers verify whether a given reply came from a specified AI model, and said its efficiency is significantly higher than previous open-source methods. The proof can be used to make it harder for attackers to hijack conversations, and can also let security evaluators confirm that the model being tested is the same as the model serving the public. Anthropic has open-sourced the FlashProve code and listed collaborators including Alessandro Chiesa.

• Anthropic releases FlashProve, using cryptography to prove the source of model inferenceUnite.AI
• Anthropic open-sources FlashProve, which can verify on an ordinary CPU which model a reply came fromRohan Paul
• Anthropic develops the FlashProve system, using cryptography to prove AI model identityTechCrunch AI

Tomasz Tunguz analyzes Anthropic S-1/A: how rising short-term GPU rental prices push up computing costs

Citing Anthropic's updated S-1/A, Tomasz Tunguz notes that short-term burst compute is priced at $7 to $13 per GPU hour, driving Anthropic's blended GPU-hour cost up 82% within two quarters, turning $974 million in Q2 operating profit into a $567 million loss in Q3. The article says long-term contract prices are about $2 to $4 per hour, and once long-term leased capacity is exhausted, buyers can only turn to the short-term rental market, while data center project delays and power shortages make supply pressure hard to ease in the short term. The author argues that the key in 2027 is to match customer contract terms with the terms of the underlying GPU leases.