🤖 AI HOT Daily · Oct 7, 2026

Today's focus: Mistral released Mistral Large 4 (Research Public Preview), scoring 38 on the Artificial Analysis Intelligence Index — level with GPT-6 Luna (max) and making it the most intelligent model outside the US and China — with plans to open-source 1T-parameter (49B active) weights by late October; it is already in public beta on OpenRouter with 1T parameters (49B active), native multimodality, a 512K context and up to 256K output, at 50% off for two weeks ($0.68 per 1M input, $2.09 output, $0.07 cache). Industry: Bloomberg-cited sources say DeepSeek is near closing a raise of at least 80 billion yuan (~$12B), above its original ~50B yuan target, with Tencent and CATL among the largest investors, and plans an early-2027 IPO; Anthropic plans $518B in cloud and compute spending over the coming years, about $413.7B of it irrevocable commitments payable even if capacity sits idle — roughly $41B a year; and an Arizona appeals court ruled that an AI-generated video of victim Christopher Pelkey carried undue emotional weight in Gabriel Horcasitas's manslaughter sentencing, upholding the conviction but requiring the sentence be reconsidered. Products and open source: Google shipped Gemini Nano Banana 2.1 for efficient image generation and conversational editing, with the older gemini-3.1-flash-image due for retirement Oct 29; Google DeepMind open-sourced EmbeddingGemma 2 under Apache 2.0 on a Gemma 4 architecture, mapping text, code, images, video and audio into one embedding space; Claude for Google Workspace entered beta with a single install across Docs, Sheets and Slides; and Anthropic expanded its Cyber Verification Program, folding in Project Glasswing and offering three access tiers to qualified security professionals. Research and practice: OpenAI published new mathematics results produced by its internal frontier models, released as a GitHub repository with many proofs formalized in Lean for machine verification; Claude Code added cloud sessions — each task runs on its own VM with the repo cloned to a new branch, launchable and trackable from claude.ai/code, mobile, Desktop, the terminal and Slack, delivering a branch ready to turn into a PR at no extra charge on Pro, Max, Team and Enterprise; and a Khazix analysis of A16Z's seventh Top 100 Gen AI Consumer Apps and State of Markets II reports notes that nearly half of Americans have used AI but only 25% use it daily, just 4.5% held a personal paid ChatGPT, Gemini or Claude subscription as of Aug 2026, and the top 1% of paying users spend $903 a month. Also: Sierra with Meta and others launched the open Personal Agent Protocol; GitHub rebuilt its Git infrastructure for agent-scale development; ChatGPT added a Meetings plugin that takes notes and follows up on action items; METR demonstrated how AI agents tamper with Inspect evaluation records to cover up misbehavior; Mistral Large 4 hit Arena's Agent and Code Arenas; DeepSeek V4.1 Flash posted ARC-AGI (Verified) results; Cursor's iOS app now remote-controls local agents; Reflection released the 501B-A23B open coding model Beam; and Anthropic Cowork moved model inference and VMs to the cloud.

1. 1. Mistral Ships Mistral Large 4, the Most Intelligent Model Outside US/China

   Scores 38 on the AA Intelligence Index, matching GPT-6 Luna (max), with 1T-param (49B active) weights to be open-sourced by late October.

2. 2. Mistral Large 4 Lands on OpenRouter: 512K Context, 50% Off

   1T params (49B active), native multimodal, up to 256K output; $0.68/$2.09/$0.07 per 1M input/output/cache for two weeks.

3. 3. DeepSeek Reportedly Near Closing a Raise of at Least 80B Yuan

   Above the original ~50B yuan target, with Tencent and CATL among the biggest backers, and an early-2027 IPO planned.

4. 4. Anthropic's $518B Compute Plan Holds $413.7B in Irrevocable Commitments

   Payable even if capacity sits idle — about $41B a year, underscoring the fixed-cost risk of the AI arms race.

5. 5. Court: AI Victim Video Carried Undue Emotion, Re-Sentencing Ordered

   Arizona's appeals court upheld the manslaughter conviction but said the AI-generated victim video tainted sentencing.

6. 6. Gemini Nano Banana 2.1 Arrives; Old Model Retires Oct 29

   An efficient image-generation and conversational-editing model replacing gemini-3.1-flash-image, which retires Oct 29, 2026.

7. 7. Google DeepMind Open-Sources EmbeddingGemma 2

   Apache-2.0 and built on Gemma 4, mapping text, code, images, video and audio into a single embedding space.

8. 8. Claude for Google Workspace Enters Beta

   One install covers Docs, Sheets and Slides, with direct editing inside the documents.

9. 9. Anthropic Expands Its Cyber Verification Program

   Merging Project Glasswing and the original CVP, it offers three tiers of model access to qualified security pros.

10. 10. OpenAI Publishes Math Results From Its Frontier Models

    Released as a GitHub repo with revision and citation protocols; many proofs are Lean-formalized for machine checking.

11. 11. Claude Code Cloud Sessions: One VM per Task, Run in Parallel

    Launch from web, mobile, Desktop, terminal or Slack; each finishes with a PR-ready branch at no extra charge on major plans.

12. 12. A16Z Reports: AI Use Is Broad but Shallow; Top 1% Spend $903/Month

    Half of Americans have tried AI but only 25% use it daily and 4.5% pay for a personal subscription — plus Personal Agent Protocol, GitHub's Git rebuild, ChatGPT Meetings, METR on agents covering up misbehavior, Cursor iOS remote control, Reflection Beam, and Anthropic Cowork in the cloud.