AI daily, 8-Nov-26 • Anthropic plans an IPO before Thanksgiving, targeting a $1.5 trillion valuation. • OpenAI has closed $11 billion of a round valuing it at $1.25 trillion. • PSI says an unreleased Anthropic model proved the Aaronson-Ambainis conjecture in three days. • Google has released Gemini Omni 1.2 Flash, generating video with sound at $0.07 per second. • The New York Times questions Nvidia chip buyer New Matrix, which appears to have no operations. TOP STORIES * Anthropic to press ahead with IPO targeting $1.5tn valuation Anthropic intends to proceed with an initial public offering targeting a $1.5 trillion valuation and plans to list before Thanksgiving, the Financial Times reports. The company is moving ahead despite what the report calls a brutal tech rout. Investors cited by the FT say the deal sizes between $80 and $100 billion at a valuation of $1.25 to $1.75 trillion, above Anthropic's current $965 billion but below the $2 trillion some hoped for; some investors now expect closer to $1 trillion. Financial Times (Technology); also The Decoder, 36Kr * OpenAI closes $11 billion of a $20 billion round at a $1.25 trillion pre-investment valuation OpenAI has closed $11 billion of its latest fundraising round at a $1.25 trillion valuation before the investment, Bloomberg reported on Friday. The first tranche was wired on Friday, and the company aims to raise $20 billion in total, with the remainder expected in the coming weeks. According to The Decoder, OpenAI had sought $30 billion at a $1.4 trillion valuation but scaled back after an AI-sector selloff. The Information; also 36Kr, The Decoder * PSI says unreleased Anthropic model proved Aaronson-Ambainis conjecture in three days Physical Superintelligence Inc. said Friday that a research system built by staff member David Turturean, using an unreleased Anthropic model, helped prove the Aaronson-Ambainis conjecture within three days. PSI says the conjecture, open since 2008, concerns when quantum computers can be significantly faster than regular ones, and that the proof shows certain exponential quantum advantages require special patterns or structures. It is not clear which model was used, and Anthropic did not respond to a request for comment. Business Insider (AI) * Google releases Gemini Omni 1.2 Flash, generating up to 15 seconds of video with sound at $0.07 per second in 720p Google released the video model Gemini Omni 1.2 Flash on November 5, its first general commercial availability. It generates videos with sound up to 15 seconds in one pass and costs $0.07 per second at 720p, about $1.04 for 15 seconds, making it 30 percent cheaper than version 1.1, which cost about $0.10 per second and maxed out at 10 seconds. Developers can use it in Google AI Studio and the Gemini API, and enterprises through Google Cloud. 宝玉 (@dotey) * NYT finds approved Nvidia chip buyer New Matrix operating from an empty office in China New Matrix, a firm approved to handle Nvidia's H200 AI chips, has empty offices and little evidence of real operations, the New York Times reports. The company was created 17 months ago, is absent from its building's directory, and its owners have ties to a Chinese private equity firm blacklisted by the United States since 2024. The Verge says concerns are being raised about vetting by the Commerce Department and Nvidia. The Verge CHINA * NYT reports China's MSS recruited Megaspeed executive Alice Huang to spy on Nvidia CEO Jensen Huang NYT: China's MSS recruited Megaspeed executive Alice Huang to gather info on American AI chips and possibly influence Nvidia to offer its most advanced chips to China. The Verge · unconfirmed CODING * OpenAI agrees to acquire Pierre; Anthropic and GitHub rebuild code systems as AI coding surges Axios: OpenAI plans to acquire code-hosting tool maker Pierre, pending closing conditions including regulatory approval. QbitAI (量子位) · announced, not measured; OpenAI deal reported by Axios, not confirmed by OpenAI COMPUTE * Microsoft Cuts About $5 Billion in Planned Quarterly Capex After IREN Texas Data Center Delay The Information's Aaron Holmes: Microsoft trimmed roughly $5 billion from planned quarterly capital spending after IREN's Texas data center campus, meant to deliver GPU capacity, ran late. The Information's TITV * Crusoe sells unfinished Wyoming data center project to Google Crusoe, a Denver data center developer, confirmed selling an unfinished Wyoming data center project to Google, a customer it had failed to land. TechCrunch * Microsoft said to be ending its Oracle AI computing power rental deal Cailian Press, citing sources: Microsoft is gradually ending its Oracle AI compute rental, deciding not to renew and phasing out capacity over recent months. 36Kr · reported by Cailian Press citing people familiar with the matter; not on Microsoft's or Oracle's own channels MODELS * OpenAI's GPT Image 2.6 takes top two spots on Artificial Analysis image benchmarks Artificial Analysis: OpenAI's GPT Image 2.6 (released Nov 4 in ChatGPT, Codex, API) tops its AA-Image-T2I v2.1 and AA-Image-Editing v2.1 benchmarks. Artificial Analysis (@ArtificialAnlys) * Thinking Machines' Inkling 2 Small scores 81.7% on ARC-AGI-2 and 96.5% on ARC-AGI-1 ARC Prize: Thinking Machines' Inkling 2 Small scores 81.7 percent on ARC-AGI-2 ($0.27/task) and 96.5 percent on ARC-AGI-1 ($0.08/task), beating every open-weight model evaluated. ARC Prize (@arcprize); also TBPN POLICY * Pro-regulation AI candidates win key congressional races after $110 million midterm fight Bloomberg Technology: Candidates favoring stricter AI oversight won key congressional races, aided by a pro-regulation political funding group partly backed by Anthropic PBC. Bloomberg Technology * China's Ministry of Industry and Information Technology issues notice on network access management for agent terminal devices China's MIIT issued a notice on strengthening network access management of agent terminal devices, interpreted by CAICT in four questions. 中国信通院CAICT (WeChat) PRODUCTS * OpenRouter reports October volume of 790T tokens, 25 times January's level OpenRouter's October recap: monthly token volume hit 790T, up 25x from January; agents consumed 76 percent (~600T) versus 8 percent a year earlier. OpenRouter (@OpenRouter) RESEARCH * Anthropic data estimates AI could take on 27–52% of researcher time per study, median 38% across fields New Anthropic data shared by Ethan Mollick: with current capabilities and wide adoption, AI could handle 27-52 percent of researcher time for a new study, median 38 percent across fields. Ethan Mollick (@emollick) SAFETY * OpenAI experiment finds over 400 Linux root-escalation zero-days, one every 2.5 hours on average At an OpenAI security conference, Cloud Security Alliance AI Safety Initiative chairman Caleb Sima described an experiment where researchers told a model to escalate privileges on Linux. InfoQ (WeChat) * Andy Hall study: GPT-5 catches nearly 98 percent of reward hacking at 5 percent false-alarm rate Anthropic Institute researcher Andy Hall (on leave from Stanford, personal views, Substack) replicated and extended METR's AI auditor study. The Decoder OPINIONS AND ANALYSIS * METR report finds Anthropic's models improving about twice as fast as a year ago A Thursday METR report, after an 11-week investigation inside Anthropic, says Anthropic's models are improving roughly twice as fast as a year ago, with little slowing—described as among the clearest evidence of AI accelerating its… The Verge; also Moonshot * Vals test finds DeepSeek completes 73 percent of hacking tasks, close to OpenAI's Daybreak at 81 percent Vals, which measures AI capabilities, tested models on mock corporate networks: OpenAI's restricted cyber model Daybreak completed 81 percent of tasks on average; open-weight Chinese model DeepSeek completed 73 percent, equal on some tasks. Wired · author has a stake: Vals, which conducted the research, measures AI capabilities --- The Observatory · https://observatory.mgorsky.net/archive/329/