AI News Today: Gemini 4 Carbon Tops Charts (Nov 5, 2026)

Google's Gemini 4 Carbon has taken the top spot on the Artificial Analysis Intelligence Index v5, one point ahead of Anthropic's Claude Opus 5.5. That is the main AI story this week, alongside a Texas lawsuit against OpenAI and new reporting on whether AI models can detect their own internal states.

Gemini 4 Carbon leads the Artificial Analysis index

Artificial Analysis reported on October 28 that Gemini 4 Carbon scores 59 on its Intelligence Index v5, ahead of Claude Opus 5.5 at 58. Carbon ranks first in four of the index's nine evaluations (HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA). Opus 5.5 still leads Artificial Analysis's Coding Agent Index, where Claude Code with Opus 5.5 scores 62 against 59 for Carbon.

On the AA-Omniscience hallucination measure, Carbon's rate is 11%, the second lowest of any model tested (only Gemini 4 Argon is lower, at 4%), compared with 50% for Opus 5.5. According to Artificial Analysis, Carbon is rolling out to selected users and is not yet generally available. One index is one yardstick, and leaderboards change quickly.

Chinese open-weight models in the rankings

Business Insider and other outlets reported that Chinese AI models trail Anthropic and OpenAI in Artificial Analysis's new rankings. The v5 index was released on October 27, and private test sets make up about 60% of it, so part of the gap reflects the redesigned benchmark.

Moonshot AI launched Kimi K3 on July 16, claiming it closes the gap with leading US models, and its weights were publicly released within days. A Lawfare analysis published on November 4 argues that open weights do not abolish power but relocate it: it shifts to whoever controls compute, hosting, standards, talent and law. The authors say governments reduce dependence mainly by diversifying suppliers across the AI stack, and that past restrictions on open technology often backfired.

Texas sues OpenAI over ChatGPT and child safety

Texas Attorney General Ken Paxton filed a lawsuit on October 23, 2026 in Collin County District Court against OpenAI and CEO Sam Altman, as The Daily Texan reported on November 2. The suit alleges violations of the Texas Deceptive Trade Practices-Consumer Protection Act, including failure to disclose ChatGPT's risks and collecting children's data without parental oversight. It asks the court to bar minors from using ChatGPT, to require a court-appointed independent evaluator to approve new ChatGPT models before release, and to award damages to Texas.

An OpenAI spokesperson said the company is willing to work with the Texas government on stronger safeguards. The Daily Texan also covers how the University of Texas is handling AI on campus, including a survey of the community on generative AI.

Can AI models detect their own internal states?

Quanta Magazine published a piece on November 4 about research into whether language models can introspect. In an earlier experiment led by Anthropic's Jack Lindsey, concepts such as "denim" were injected into a model's activity. In about 20% of trials with the most capable Claude version, the model flagged the injection, and control runs produced no false positives.

Follow-up work by Kyle Mahowald and Harvey Lederman suggests this is closer to "anomaly detection": the model senses that something unusual is happening internally but cannot identify the specifics. Whether that counts as introspection is still debated.

Why it matters

The lead at the top of the AI race is narrow and depends on the index you use (Gemini 4 Carbon versus Claude Opus 5.5). Legal pressure on AI companies over child safety is growing, and researchers are still debating how much AI models can tell about their own internal workings.