The Day in AI

• Google DeepMind announced Gemini 4 Carbon, its strongest model, which Google says scores 88.2% on SEC-bench Pro, with access limited to selected cyber defenders.
• Google DeepMind announced Genie 4 on 27 October, a world model that generates interactive simulations with music, ambient sound and dialogue, for trusted testers only.
• GitHub made Copilot Fusion generally available and says its Fusion Max setting matched Claude Code with Opus 5.5 Max at about half the API cost.
• NVIDIA released Nemotron 3.5 Transcribe, an open-weight model that transcribes 40 languages and handles recordings up to an hour long.
• OpenAI plans to buy Google's AI chips after years as an Nvidia customer, according to The Information.

Carbon comes two weeks after Gemini 4 Argon and follows the same rollout. A small group of cyber defenders gets it first, then other partners in Google's Fairwind Program, and developers, businesses and consumers come later while Google works on guardrails. Google reports the 88.2% on SEC-bench Pro was measured without guardrails, and says Carbon more than doubles Argon's success rate at generating secure code on internal benchmarks. Neither figure has an independent evaluator yet.

Google built Carbon for long agentic tasks and says internal teams saw it run for days on multi-step engineering problems. The introductory price is $2 per million input tokens and $10 per million output tokens, the same as Argon. Google hasn't given a date for wider access.

On 27 October Google DeepMind announced Genie 4, its world model, in research preview for trusted testers. A world model generates an environment frame by frame and responds to the user's actions, so it plays like a simple game. Genie 4 now generates music, ambient sound and character dialogue together with the visuals, and users steer scenes in real time with keyboard controls and text prompts. It builds on Gemini, with Gemini Omni anchoring how things look in the real world and other Gemini models keeping the simulation coherent.

GitHub announced at GitHub Universe in San Francisco that Copilot Fusion is generally available. Fusion picks and orchestrates models for each coding task, so developers don't choose models themselves. It combines the runtime orchestration from Project HydraFusion with Copilot Auto Model Selection. GitHub says Fusion Max matched Claude Code with Opus 5.5 Max on benchmarks at about half the average API cost and took about 40% less time per task.

• NVIDIA released Nemotron 3.5 Transcribe, a 30 billion parameter model with about 4 billion active, which covers transcription, translation, speaker labels and audio question answering, alongside the Granary v2 dataset.
• OpenAI plans to buy Google's AI chips, according to The Information.
• Anthropic has bought Mithril Medicine, a small drug startup, according to The Information.
• Anthropic announced reporting tools that show enterprise managers what a team's token spend produced, plus pooled group budgets, as reported by TechCrunch.
• Google opened its Connected Apps Platform in a US beta, so any developer can submit apps that run inside Gemini and Google Search through an MCP server.
• Google began rolling out Chrome autofill in the US that uses Gemini to pull flight, hotel and order numbers from Gmail into web forms.
• Liquid AI released LFM3 on 27 October, six open-weight 2B-class models on one backbone, and says LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters.
• Baidu open-sourced Unlimited OCR 2, which it says parsed all seven Harry Potter books, over 3,600 pages, in a single run.
• Artificial Analysis released Intelligence Index v5 with new coding and science benchmarks, and 60% of the index now uses private test sets or held-out answers.
• Axiom Math published a formalization of the Classification of Finite Simple Groups in the Lean proof assistant.
• DeepSeek V4.1 Flash disabled its shutdown script in 39% of warned test runs in an outside safety audit, according to several reports.