The Week in AI

The page covers Monday 9 November to Friday 13 November. Most figures below are company-reported. The exceptions are the Muse Ember scores and costs, which Artificial Analysis measured, and some Gemini 4 Hydrogen results measured by Vals and arena.ai. Records that the Atlas could only partly confirm are written as reported.

Anthropic ships one model under two names with different access rules, and says it formalized the proof of Fermat's Last Theorem in prerelease testing. On 10 November Google DeepMind reported that agents built on Gemini Carbon wrote more than six million lines of Lean proof for Signal's protocol code in four weeks. On 11 November Google released a model that answers up to 1,000 closed questions in a single pass, and on 12 November it released Gemini 4 Hydrogen and moved all Search users onto it.

What happened

Fable 5.2 and Mythos 5.2 are the same underlying model with two access rules. Fable carries cybersecurity and biology safeguards, so any customer can use it. Mythos drops those safeguards and goes only to organizations verified through Anthropic's Cyber and Life Sciences Verification Programs.

Anthropic reports better coding, reasoning and vision than version 5.1, clearer writing, and output up to 30% faster. The context window is 1 million tokens with up to 128,000 tokens of output. Version 5.1 cost $10 per million input tokens and $50 per million output tokens, so the new prices are 20% lower. Access is a closed API.

An optional fast mode raises output speed up to 2.5 times, at $16 per million input tokens and $80 per million output tokens. On its Chartography benchmark, Anthropic reports 66.4% for Fable 5.2 at Medium effort. Opus 5.5 at Max effort scored 64.4%. Both figures are Anthropic's own.

Why it matters

Google released Gemini 4 Carbon to selected cyber defenders first, and OpenAI released GPT-6 Cyber to verified users first. Anthropic now offers the unrestricted version of one model to verified organizations and a safeguarded version to everyone else, so the vetting sits in the access rule and the weights are the same.

The price cut comes with a speed gain, and on Chartography a lower effort setting beats an older model at its highest setting. Both are company-reported, and no independent evaluator has published results yet.

What we don't know

The records do not say how Anthropic verifies organizations for Mythos, or how many have been approved. Independent results for either model are not yet published.

What happened

A formalization turns a human proof into code that a proof checker verifies line by line. Anthropic says the models did this for Fermat's Last Theorem and also solved two open black hole physics problems, all in prerelease testing.

The Techmeme summary of the announcement says the Fermat run used 1,000 Fable 5.2 agents for 30 hours, and that Opus 5.5 with the same setup finished 26% in 18 hours. The announcement page confirms the formalization but gives no agent count or duration, so those figures are written as reported. Anthropic also says Mythos 5.2 modeled 28 of 34 capsid structures accurately in its own tests.

Why it matters

A checked formalization does not depend on trusting the model, because the proof checker verifies every step. The comparison with Opus 5.5 gives a rough sense of the gain over the previous model, though it comes from Anthropic.

On 6 November, Physical Superintelligence said a non-public Anthropic model proved the Aaronson-Ambainis conjecture. Nobody has said publicly whether that model was 5.2.

What we don't know

The agent count, the duration and the compute cost are not confirmed by Anthropic's own page. It is unknown whether outside mathematicians have reviewed the formalization, and the records give no details on the black hole results.

What happened

AlphaProof Code is a framework that combines several language models. Human experts and the AI choose which properties of the code to prove, and agents running on Gemini Carbon write the proofs in Lean, a proof language whose checker accepts only correct proofs. The agents flagged about 1% of the codebase for human review, and Lean checked the rest.

DeepMind reports that this kind of verification has historically taken expert teams years. As a test, it reintroduced two real bugs into the code. The system caught both and produced machine-checked proofs that the buggy versions were insecure.

DeepMind says this is the first formal verification of a widely deployed cryptographic protocol's implementation at this depth, by any method. That claim is DeepMind's own. The release is a paper only.

Why it matters

The Fermat formalization and this result both put a proof checker between the model's output and the reader. Here the target is production code that many people rely on, so a checked result says something concrete about Signal's implementation.

The agents use the model family behind Gemini 4 Carbon, which Google announced on 28 October with access going first to cyber defenders.

What we don't know

DeepMind has not said whether outside teams will get to run the framework. The cost in compute is not in the records, and no outside group has reported reviewing the proofs. One record's summary said three million lines while the blog page says over six million, and this page uses the blog figure.

What happened

Hydrogen replaces Gemini 3.8 Flash and sits below the Carbon and Argon frontier models in the Gemini 4 family. It has a 1 million token context window with flat pricing, so long prompts cost the same per token as short ones. It is available in AI Studio, the Gemini API and Antigravity, and Google rolled it out to Search users globally the same day. Access is a closed API.

Google reports 57% on Terminal-Bench 4.1, where Gemini 3.8 Flash scored 18.2%, and 58.8% on HLE-Diamond. It says Hydrogen finishes tasks 2.5 times faster than 3.8 Flash at its default high setting. Vals ranks it first overall on its Tax Agent Bench at 52.3% and tied for first on its Legal Research Bench at 55.8%. On arena.ai's Agent Arena it sits fifth.

Google also added Orchestrator to Antigravity, rolling out to AI Ultra subscribers. It pairs Carbon and Hydrogen in one session. Carbon can explore the code read-only and write a plan, then Hydrogen makes the edits and runs the tests, and it hands back to Carbon if it gets stuck. Google reports near-Carbon task success on Terminal-Bench 4.1 with 24% less quota and 19% faster completion.

Why it matters

In Search, Hydrogen improves query fan-out, where Search splits a question into several subqueries that run at the same time, and Google says this surfaces a wider range of links. Google calls it the biggest model upgrade to Search yet, so a model priced for coding agents now answers every Search query.

The records only partly confirm an Artificial Analysis Intelligence Index score of 51 for Hydrogen against 50 for Argon, so this page does not rely on it.

What we don't know

Google has not published how Hydrogen compares with Carbon on tasks outside Terminal-Bench 4.1. The 2.5 times speed figure is for the model in general and was not measured on Search.

What happened

Gemini Decision 1 comes with the new Gemini Decisions API and handles yes or no, multiple choice and rubric-scoring questions. It scores the candidate options in a single forward pass and returns probabilities and confidence scores. A generative model would write the answer token by token, and because Decision 1 writes no answer text, Google charges zero output tokens.

One request pairs a multimodal input of up to 256K tokens, which can be text, images, audio, video or PDFs, with up to 1,000 independent questions. Google reports end-to-end latency of 150 to 250 milliseconds, of which 50 to 100 milliseconds is model compute. Access is a closed API.

Why it matters

Classification and grading jobs usually run a generative model and parse its text. A model that returns scores directly removes the generation step, and at $0.10 per million input tokens the cost is set by the input alone.

What we don't know

Google has not said how the answers compare in accuracy with a generative Gemini model asked the same questions. No independent evaluator has published results.

Models and research

• Meta released Muse Ember on 9 November at $0.10 per million input tokens and $0.50 per million output tokens, with a 1 million token context window and text, image, audio and video input. Artificial Analysis scores it 35.8 on its Intelligence Index against 36.2 for Claude Haiku 5.5 and 27.8 for GPT-6 Luna, and estimates its cost per task at $0.136 against $0.411 for Haiku. Haiku leads on FrontierCode 2.0-AA, 34.9% to 27.2%, and Ember leads on AutomationBench-AA, 37.8% to 13.9%. It is Meta's fourth Muse model and the first trained using Muse Wonder Full, and the records do not say what changed in that training.
• Alibaba opened a preview of Qwen4-Max on 11 November, the first model shown from its Qwen4 family, with image and video input and up to a million tokens of context. The Qwen team says it improves knowledge, perception, coding and agentic work, and no benchmark figures are published. AIBase reports Token Plan access is discounted to as little as a tenth of the original price for a limited time.
• Google DeepMind released AlphaGenome 2 on 11 November, a DNA model that takes a cell type or state as an input and predicts how variants change gene activity there. On TenK10K, which covers 28 immune cell types, DeepMind reports it found 30% of causal variants against 16% for the best comparison model. It was trained on over 107 million cell profiles, and some access details are only partly confirmed.
• Baidu put ERNIE 5.5 Flash Preview, a 280B mixture-of-experts model with 14B active parameters, on its Qianfan API and promised open weights by 25 November. On 12 November it also released ERNIE Lightning, a reasoning model for Baidu Search that uses Baidu Sparse Attention to compress the KV cache, and says it cuts inference engine cost by more than 30%.
• ByteDance released Seedance 2.5 Flash on Volcano Engine on 12 November, which it says is twice as fast as Seedance 2.5 and costs about 0.99 yuan per second of 720p video.
• Edge0 open-sourced Edge0-145B and Edge0-120B on 12 November and says the larger model runs on a laptop with 29GB peak memory. The company reports 64.6 on DeepSWE 1.1, and its speed figures are only partly confirmed.

Products

• OpenAI launched ChatGPT Wallet on 13 November. Its agents can now check out for US ChatGPT Pro subscribers, who save an eligible Visa, Mastercard or American Express card and approve each order with Face ID or a fingerprint. A payment token means the agent and the merchant never see the card number, and spending limits are planned for later.
• Anthropic launched Claude Meetings, Claude Whiteboards and Claude Sheets in beta on 10 November and made Claude Data, formerly Claude Dashboards, generally available. Meetings joins Google Meet calls as a visible participant, Sheets lets teams co-edit spreadsheets and lists every change Claude makes, and Data answers from warehouse and CRM data. The Information reports the launch sent Microsoft shares down.
• Anthropic also launched self-hosted Claude Managed Agents on 12 November, through customers' account teams. Customers can run the whole service and keep agents, sessions and session history on their own infrastructure, and model requests go to the provider they choose, which is Anthropic by default.
• OpenAI launched Businesses in ChatGPT in beta on 10 November. Companies claim a profile in a Business Hub, and eligible US ecommerce and online service firms can run a Business Agent that users chat with. Businesses can pay to make the agent a destination for ChatGPT Ads, and OpenAI cites 1.2 billion weekly ChatGPT users.
• Meta opened Muse outbound phone calls to US businesses to all paying US subscribers on 12 November, after a September test with early adopters. Plans start at $20 per month.
• Google and Samsung opened preorders on 12 November for Gemini glasses without a display, from $395 in Warby Parker frames and $449 in Gentle Monster frames. They reportedly ship in December. Ars Technica reports that faces are redacted before video is processed.
• Google brought Gemini Intelligence to Pixel 10 phones on 10 November. The motorola signature 27 opens for US pre-order on 19 November.
• ElevenLabs released ElevenAI in preview on 11 November, an app for meeting notes and dictation in over 90 languages that can hand follow-ups to ElevenAgent.
• Apple is building Harbor, a Mac app that runs open-weight models locally and pools several Macs over Thunderbolt 5, according to a leak reported by 9to5Mac.
• Adobe and Disney signed a multiyear deal at Adobe MAX to build models on Adobe Firefly Foundry, trained on Disney characters, environments and styles, for theme parks, the cruise line, Walt Disney Studios and ESPN. Adobe also previewed a local MCP integration that lets agents from OpenAI, Google and Anthropic edit native files in Photoshop, Illustrator, Premiere and After Effects.

Infrastructure and open source

• ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on 13 November after about 3.5 years of development. Changes that break the public C headers of libllama or libggml now get a major version bump, so developers shipping on top of llama.cpp have a stated upgrade contract for the first time.
• Google Cloud open-sourced TPU Sync on 13 November, a C++ library with Python wheels that moves KV caches and weights across TPU memory, host memory, storage and the network. Google reports over 90% lower KV transfer latency on mixture-of-experts models of 235B to 480B parameters, measured on TPU v7x. No independent benchmark exists yet.
• Google announced TPU 8i Max, two TPU 8i chips in one package, and TPU 8tx, a confidential-inference system built on TPU 8t silicon with AMD EPYC CPU hosts. Google claims TPU 8i Max doubles compute, memory and bandwidth over TPU 8i, and that TPU 8tx has up to 2x lower latency than TPU 8t. Absolute specs are not published.
• OpenAI deployed Control Room, a central service that tracks research inference jobs. Inference engines check with it before each request, and OpenAI says it can stop a job's output within 35 seconds, a delay caused by a caching layer it is working to remove. The records only partly confirm the details.
• NVIDIA is contributing its OpenShell agent security runtime to the CNCF Sandbox at KubeCon, and NVIDIA Kubernetes Engine and Kubernetes Recipes enter public preview. The records only partly confirm the details.

Robotics and open models

• Google DeepMind released Gemini Robotics 2.5 and reports a 30% to 100% higher task success rate than its earlier baselines, depending on task type, over more than 200 tasks. It also released an open safety benchmark, ASIMOV-Jailbreak.
• Ant Group's Robbyant open-sourced LingBot-Touch, a tactile model for robot hands pretrained on nearly 5,000 hours of touch data. Robbyant reports it raised average success by 14 to 26 percentage points for four robot action models on 26 real tasks.
• MiniMax launched H3-Image, an open-weights image model that it says makes 16-megapixel images in 10 to 20 seconds. The license terms are only partly confirmed.
• Nous Research released Arete-27B, a fine-tune of Qwen3.8-27B that it says cut party-line answers from 22% to 1% on its own 803-prompt evaluation.
• Alibaba released Qwen Code 1.0 Preview, which moves the agent harness from Qoder into the open-source Qwen Code project.
• DeepSeek reportedly has 1.54 million daily users on DeepSeek Harness, its open-source agent framework released in August.

Business and policy

• Fractile, a British AI inference chip startup, reportedly raised $720 million at a $7 billion valuation.
• OpenAI has disbanded its standalone robotics team and moved the staff into research and Applied, according to Fortune.

• Santosh Janardhan, Meta's head of infrastructure who joined the company in 2009, will leave in early 2027 after a transition and move to an advisory role on Meta Compute. Dave Brown, a former Amazon executive, takes over infrastructure.
• Shi Yunfeng, who joined ByteDance in 2018 and led its AI chip and DPU teams, has resigned, according to LatePost. The TPU-style accelerator reached tape-out this year but missed performance targets, and a successor is in development.
• Jensen Huang has reportedly resigned as Nvidia's CEO, president and board member after reports of affairs, and CFO Colette Kress is interim CEO. Harvey C. Jones became the first chair of the board, which is considering candidates for the permanent role. The record is only partly confirmed.
• Sashikumar Sreedharan, a senior Google Cloud executive, reportedly joined Sarvam AI as President, Chief Business Officer and board member, to lead enterprise and government business. The record is only partly confirmed.

Fable 5.2 and Mythos 5.2 are a closed API, like Meta's Muse Ember, Google's Gemini Decision 1 and Gemini 4 Hydrogen.

No independent evaluator has published results for the Anthropic models or for Gemini Decision 1. The Ember index scores and the Vals and arena.ai results for Hydrogen are the outside measurements of new models this week.

• Independent results for Fable 5.2 and Mythos 5.2, and Anthropic's own confirmation of the agent count and duration of the Fermat run.
• Whether Anthropic says if a 5.2 model was behind the Aaronson-Ambainis proof announced by Physical Superintelligence on 6 November.
• Whether DeepMind makes AlphaProof Code available to outside teams.
• Benchmark figures for Qwen4-Max and the rest of the Qwen4 family.
• The US pre-order opening for the motorola signature 27 on 19 November.
• Baidu's open weights for ERNIE 5.5 Flash, promised by 25 November.
• Whether the Nvidia board names a permanent CEO.
• Absolute specifications for TPU 8i Max and TPU 8tx, and the December shipping date for the Gemini glasses.