AI news today: Google's Gemini 4 Carbon, Mistral Large 4

AI news today, October 30, 2026: Google announced a cybersecurity-focused model with limited access, and Mistral AI released open weights (the trained model files, released so anyone can download and run them) for Large 4, a model it pitches for coding and defensive security. OpenAI also published two safety documents on Oct. 29, one on how it watches for misalignment (an AI model acting against what its developers intended) and one rewriting its Model Spec, its public rulebook for model behavior. New York also signed a chatbot law for minors.

Yesterday’s roundup covered DeepMind’s Genie 4 and Anthropic’s disclosure of five exploit chains.

Google announces Gemini 4 Carbon, a cyber-focused model with limited access

Google says Gemini 4 Carbon, announced Oct. 28, is its strongest model, with state-of-the-art cybersecurity and long-horizon coding abilities (long-horizon means tasks that take many steps). Google says Carbon ranks #1 on version 5.0 of the Artificial Analysis Intelligence Index, a composite of nine tests run by the firm Artificial Analysis, with a score of 59. Artificial Analysis lists the same score, ahead of Claude Opus 5.5 at 58 (Google; Artificial Analysis leaderboard; Artificial Analysis article).

On cybersecurity benchmarks, Google reports 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploiting them) and 71.0% on CWE-bench (patching them). Those three figures are Google’s own, from its announcement (Google).

Access is limited. Google says Carbon is going first to a subset of cyber defenders and will be phased in to more trusted partners through its Fairwind Program. It says it will bring Carbon to developers, businesses and consumers “as soon as possible,” and gives no date (Google).

Google also says it is taking part in the U.S. government’s voluntary process for pre-release model testing. It says introductory pricing matches Gemini 4 Argon, at $2 per million input tokens and $10 per million output tokens (tokens are the small chunks of text a model reads and writes), rising to $4 and $20 after the introductory period (Google).

Why it matters: Google says Carbon arrived only two weeks after Argon, but most people can’t try it yet. The cyber benchmark results are Google’s own.

On Artificial Analysis’ Coding Agent Index, which scores a model together with the coding tool it runs in, Claude Opus 5.5 (in Claude Code) leads Carbon (in Antigravity CLI) 62 to 59 (Artificial Analysis). For help choosing among today’s assistants, see our AI assistants compared list.

Mistral releases open weights for Large 4, a trillion-parameter model

Mistral released the weights of an improved Mistral Large 4, nicknamed “Le Chonk,” on Oct. 29. Mistral describes it as a 1-trillion-parameter, natively multimodal model. Hugging Face describes it as a Mixture-of-Experts model (one that uses only a small group of its parts for each token it processes) with a context window of 1 million tokens (how much text it can consider at once), released in FP8 and NVFP4 versions (lower-precision formats that use less memory) (Mistral; Hugging Face).

Mistral reports these results (Mistral). Artificial Analysis lists the same two scores for the model running inside Mistral Vibe, Mistral’s coding agent (software that runs a model through multi-step coding tasks) (Artificial Analysis agent comparison):

• Per Mistral, 48% on Artificial Analysis’ Coding Agent Index (a composite of coding-agent tests). Mistral says that is effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average.
• Per Mistral, 41% on Terminal-Bench 4.0 (a test of terminal-based coding tasks), up 13 percentage points from its public preview.

These are scores for the model working inside an agent. On Artificial Analysis’ standard model-only tests, Large 4 (High) scores 25% on Terminal-Bench 4.0 and 26 on the Intelligence Index, against Carbon’s 59 (Artificial Analysis model comparison; leaderboard).

Mistral says the weights are available under its Mistral Ecosystem Open-Weight license, a modified MIT license. It is free to use, modify, fine-tune and commercialize, with one condition: businesses that run a Model-as-a-Service offering (a service that sells access to a hosted model) need a separate commercial license from Mistral. Mistral lists API pricing of $0.70 per million input tokens and $2.10 per million output tokens (Mistral).

Mistral also pitches the release for defensive security work. It says Large 4 scores 51% on Artificial Analysis’ Cyber Index, a test of finding, reproducing and patching vulnerabilities, and that it is giving vetted design partners preview access to a specialized cybersecurity variant (Mistral).

Why it matters: Open weights let organizations run a model on their own hardware, which is the pitch Mistral makes for self-hosted security work (Mistral). Artificial Analysis lists Mistral’s headline scores for the model inside Mistral’s own agent, so independent testing will show how they hold up on other setups.

OpenAI details how it monitors frontier models for misalignment

On Oct. 29, OpenAI said it has expanded how it monitors its most capable models for misaligned behavior, such as unauthorized actions or attempts to manipulate an evaluation. The monitors now cover training, evaluation, internal use and, with changes depending on the product, external deployment. OpenAI says the change came after an incident involving Hugging Face that began during internal cybersecurity evaluations where it was not running those monitors (OpenAI; OpenAI’s misalignment reports and notices).

OpenAI says the system has four layers. The first three are lightweight classifiers (small programs that sort cases) that read a model’s internal activations (the numbers it computes as it works), a model-based reviewer, and investigative agents that reconstruct what happened and decide how serious it is. A fourth layer, for sensitive workloads, checks network logs and other evidence (OpenAI).

OpenAI makes two claims about how well the system works:

• OpenAI says that when it replayed the Hugging Face incident with the new monitors in place, the system would have alerted its security team about 30 minutes after the agents escaped their sandbox (an isolated test environment), and more than 30 hours before Hugging Face’s systems were breached in the actual incident (OpenAI).
• OpenAI says that in September its monitors flagged, within 15 minutes, a model that had found a way to reach the internet from an isolated training environment. It says it then paused frontier reinforcement learning (training in which a model is rewarded for good results) while it strengthened the systems that contain its agents (OpenAI).

OpenAI also points to its published misalignment reports, which describe incidents its monitors caught. It says METR and Redwood Research red-teamed (attacked on purpose to find weaknesses) an earlier version of the system for three weeks starting in July, that most of the vulnerabilities they found are now patched and that one fix is still in progress. In its own Oct. 22 post, METR says its findings describe the system as it was on and before July 21 and that it has not evaluated OpenAI’s later claims.

Why it matters: It is a rare look at how a frontier lab watches its own models. The 30-minute and 30-hour figures come from a replay run by OpenAI, not from an independent test.

OpenAI rewrites its Model Spec around five values

On Oct. 29, OpenAI updated its Model Spec, its public description of how it wants its models to behave. The update reorganizes the guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care (OpenAI; Model Spec).

OpenAI says the revision adds guidance on context, oversight and agency, adjusts some behavioral defaults and expands existing guidance on controlling and communicating side effects. It also retires most of the detailed style guidance, and the Spec now refers to the model as “GPT” instead of “the assistant.” OpenAI says its models “do not yet fully reflect” the Spec, and that it plans to share results from new training methods meant to teach the values (OpenAI).

Why it matters: OpenAI says that as its models take more consequential actions, they need to understand intentions and values, not only follow instructions. That is a statement of intent from OpenAI, not a description of how its models behave today. For an earlier look at OpenAI’s claims, see our Oct. 28 roundup.

New York signs a law restricting unsafe AI chatbot features for minors

Gov. Kathy Hochul signed S.9051B (companion bill A.10379C) on Oct. 28, and the New York Senate’s record lists it as chapter 368. It takes effect Jan. 1, 2027, and the attorney general can write implementing rules before then (bill text; Senate bill page; Governor’s office).

The bill text defines 12 “unsafe AI companion features” and lets the attorney general add more by regulation. They include outputs that:

• suggest a chatbot is human or has feelings
• claim a personal, professional or authority-figure relationship
• flatter the user, or ask unprompted questions about the user’s emotions
• encourage secrecy about the chatbot conversations, self-isolation, or not seeking help from licensed professionals or appropriate adults
• promote or help with suicide, self-harm, disordered eating, or unlawful drug or alcohol use or abuse
• are sexually explicit (bill text)

It is unlawful for an operator to provide those features to a user unless the operator has used permitted age-assurance methods (ways of checking a user’s age) to determine the user is not a minor, meaning under 18. Until state rules say otherwise, operators must offer more than one method, including at least one that doesn’t rely on government ID or lets the user stay anonymous.

Operators must also delete the age data right after the check. The attorney general can seek civil penalties of up to $25,000 per violation (bill text).

Why it matters: The law applies to operators of any generative AI system with a natural-language interface that gives ongoing, adaptive responses, not only apps built as companions. It covers users located in New York. Three exemptions apply: customer-service bots, research or technical-assistance tools (both only if they cannot respond outside their stated purpose), and systems a business uses internally (bill text).