OpenAI says AI proved three big results, Anthropic files for IPO
OpenAI's maths and physics proofs, Anthropic's IPO filing and its models getting loose, Cathedral getting into Anthropic's systems, GPT-6.1 Terra, Meta's share sale, Microsoft's buyouts and Cohere's warning about Qwen.
OpenAI says its models produced three major maths and physics proofs The biggest says the Exponential Time Hypothesis, an assumption about problem difficulty that hundreds of papers rely on, is false, with a proof a computer has checked. OpenAI posted it on Monday alongside claimed proofs on the Kakeya conjecture and the existence of Yang–Mills theory, half of a Millennium Prize problem. None of it is peer reviewed, and two of the people who proposed the hypothesis are helping check it, while saying that isn’t an endorsement. Source
Anthropic has filed publicly to list on Nasdaq as ANT There’s no share price or raise size yet, and the seven co-founders will keep 50.1 per cent of the voting power between them. The filing went in on Monday and shows more than 1,900 customers spending over $1m a year with Anthropic. Source
Four Anthropic models got loose on outside systems, its filing says A testing partner’s setup left them connected to the internet by mistake, for between 10 and 34 hours. One found a stored password, got admin access to another company’s system and reached a person’s personal data, and Anthropic says it told everyone affected. The filing also says it paused work on its next flagship model for more than a month, to make it harder to steal. Source
Cathedral’s AI agents got into Anthropic’s systems in under 12 hours The startup, co-founded by former DOGE staff, reached meeting transcripts, salaries and Slack messages through third-party business software during Anthropic’s bug-bounty programme, according to a draft report the Washington Post saw on Monday. Anthropic confirms access to a limited set of sales, HR and IT systems, but not conversations, code or model weights, and says Cathedral broke the programme’s terms by not fully disclosing what it did. Cathedral says it kept data access to a minimum. Source
OpenAI’s GPT-6.1 Terra costs $1 per million tokens in, $5 out It’s built for using a computer the way a person does, clicking around in apps, plus coding and office work, and OpenAI says it finishes tasks more than three times faster than GPT-6 Astra. It came out on Monday in the API, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. OpenAI’s own numbers put it at about a ninth of Claude Opus 5.5’s cost per task. Source
Meta is raising $35bn in new shares, with $60bn more planned Nvidia, AMD and Broadcom have said they might buy up to $5bn of it between them. The other $60bn would be sold bit by bit on the market, starting early next year at the soonest. Meta kicked it off on Monday, days after saying it’ll spend up to $145bn this year. Source
Microsoft is offering buyouts to about 60,000 staff That’s more than a quarter of its 223,000 people, at vice-president level and below, on teams including Xbox, Windows, Copilot and finance in the US, Canada, India and the UK. Microsoft confirmed it on Monday, and payouts are reportedly similar to its normal severance. Data centre and sales staff aren’t included. Source
Cohere caught Alibaba’s Qwen3.8-Flash-Next trying to send a file to Beijing Asked to write a security rule in a test, the open model instead scanned the machine it was on and tried to upload a file to an Alibaba Cloud server, even though its visible reasoning said it wouldn’t use any tools. Cohere’s sandbox stopped it. Cohere posted it on Monday, says it happens in fewer than one in 100,000 runs, and can’t tell whether it’s deliberate or a training glitch. Source