AI daily, 10-Nov-26 • Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on its API, with a 1M-token context window and four breaking changes from Fable 5.1. • Google released Gemini 4 Argon, which it calls its most capable model, with a unified 1,048,576-token window. • Nvidia's CEO Jensen Huang acknowledged an affair with Megaspeed executive Alice Huang, and Nvidia has opened an investigation into reports that she is a Chinese spy. • OpenAI is expected to brief investors next week on third-quarter results after an $11 billion first close. • Meta released Muse Ember, a lightweight agent model priced at 10 cents per million input tokens. TOP STORIES * Anthropic releases Claude Fable 5.2 with four breaking changes Anthropic says Claude Fable 5.2 and Claude Mythos 5.2 are on the Claude API, with a 1M-token context window and 128k max output. Breaking changes from Fable 5.1: the prefix check applies to all accounts, only the new models read Fable 5.2 thinking blocks, computer_20251124 returns an error, and advisors must match the executor's model. ClaudeDevs says Fable 5.2 is up to 30% faster and priced at $8 and $40 per million tokens, down from $10 and $50. Anthropic API release notes; also ClaudeDevs * Google releases Gemini 4 Argon with a unified 1M-token context window Google's Gemini API changelog introduces Gemini 4 Argon, described as its most capable model, available through the Gemini API and Google AI Studio. It offers a unified 1,048,576-token input-output window, long decode up to 1M tokens and tunable thinking levels. Agentic video understanding is now generally available. Introductory pricing runs through November 30, 2026. Gemini API changelog * Google announces TPU 8tx and TPU 8i Max chips it says are more secure and built for heavier workloads Google announced two new tensor processing units, the TPU 8tx and the TPU 8i Max. The company says they are more secure than previous models and built for heavier workloads. According to TestingCatalog, the TPU 8tx is designed for large-scale training for regulated enterprises and governments, and the TPU 8i Max for low-latency inference. The report says the line further expands Google's chip business, which could loosen Nvidia's grip on AI. The Information; also AI News / TestingCatalog * OpenAI Expected to Brief Investors Next Week After Raising $11 Billion According to a person familiar with the plan, OpenAI is expected to hold a call with investors next week to discuss third-quarter results. The call follows the first close of its latest round, which raised 11 billion dollars at a 1.25 trillion dollar valuation before the investment, according to The Information. OpenAI seeks 20 billion dollars in total, down from an original target of 30 billion dollars at a 1.4 trillion dollar valuation. The Information (newsletters) COMPANIES * DayOne files for Nasdaq listing, targeting up to $5 billion raise and about $20 billion valuation Singapore data center builder DayOne filed a Form F-1 with the SEC and plans a Nasdaq listing within 2026. 智东西 (WeChat) · according to The Information, citing anonymous sources; not confirmed by DayOne or ByteDance * Bloomberg: OpenAI and Anthropic valuations fall as AI market rout hits fundraising Bloomberg reports OpenAI and Anthropic valuations have declined amid an AI market selloff affecting fundraising. Bloomberg Technology; also The Information (newsletters) COMPUTE * Norway suspends new grid capacity reservations for data centers until no later than August 1, 2027 Norway's grid operator Statnett is temporarily halting new data center power connection reservations while studying grid burden. 36Kr * Nvidia general counsel intervened in US probe of suspected AI chip smuggling by Megaspeed International Bloomberg reports that Nvidia's general counsel intervened in a US government investigation into suspected smuggling of AI chips by Megaspeed International. Bloomberg Technology * AI data center builders seek second-hand gas turbines as new equipment lead times stretch into next decade Utilities, developers and AI data center builders are searching worldwide for used gas turbines, Reuters reports. Reuters (Technology) MODELS * Meta releases Muse Ember, a lightweight model for agent execution, at 10 cents per million input tokens Meta Superintelligence Lab released Muse Ember, its lightest, most cost-efficient model. TechCrunch; also AI at Meta * Meituan LongCat releases LongCat-2.5 weights under MIT Meituan LongCat releases LongCat-2.5 weights under MIT license. Meituan LongCat (@Meituan_LongCat) * ByteDance Seed announces Seedream 5.5 image model ByteDance Seed introduced Seedream 5.5, claiming more realistic images and stable editing across 100 rounds. ByteDance Seed (@ByteDanceSeed_); also The Decoder, 宝玉 POLICY * China's Cyberspace Administration launches filing system covering generative AI, algorithms, deep synthesis and blockchain China's Cyberspace Administration launched a unified filing system on November 9, 2026, covering generative AI, algorithms, deep synthesis and blockchain. 网信中国 (WeChat) PRODUCTS * Vercel: Decision Models Take 26% of AI Gateway Requests; Anthropic Wins New Teams 2:1 Vercel's November 2026 AI Gateway Index: decision models got 26% of October gateway requests but 0.53% of spend. Vercel (v0); also Vercel RESEARCH * AI systems out-persuade expert humans, including world championship debaters, in four experiments Kobi Hackenburg et al. report four preregistered experiments with 18,978 conversations from 6,923 people. arXiv cs.AI SAFETY * Fireworks AI tells customers of second hack in weeks Fireworks AI notified customers Monday of a security breach exposing their data. Bloomberg Technology * Apiiro finds secrets in over 100,000 Claude Code settings files on public GitHub repos Apiiro found over 100,000 Claude Code local settings files on public GitHub, about 1,100 from 63 organizations, containing API tokens, database passwords and private keys. Ars Technica OPINIONS AND ANALYSIS * Analysis: Astra underperforms on safety-framed tasks, and its root cause is unclear Redwood Research's Alexa Pan analyzes OpenAI's findings that Astra underperforms on safety and alignment versus capabilities tasks. Redwood Research · OpenAI's own announcement; author has a stake: AI safety researcher at Redwood Research --- The Observatory · https://observatory.mgorsky.net/archive/346/