Google announced Gemini 4 Carbon on October 28, a model built for cybersecurity and long-running software work that scores 59 on the newly rebuilt Artificial Analysis Intelligence Index, one point above Anthropic's Claude Opus 5.5 and the highest published result. The release arrives just 15 days after Gemini 4 Argon reached customers, and it returns Google to the top of a closely watched ranking for the first time since a run of missed targets through 2026. Most customers cannot use it: Carbon is going first to a subset of vetted cyber defenders, with no date set for developers, businesses or consumers.
Google pitches Carbon at three security jobs, finding software flaws, checking whether they can be exploited, and repairing their root causes. It reports 71 percent on CWE-bench, a defensive-coding test where a fix must block the exploit and keep every existing test passing, three points ahead of GPT-6 Astra, Grok 4.7 and Argon on the single-attempt measure. Koray Kavukcuoglu, who has led Google DeepMind since Demis Hassabis stepped down in August, wrote that internal teams kept Carbon working on single engineering problems for days. Google did not publish completion rates for those multi-day tasks, and says its security results were measured without guardrails, so customers on the restricted version may not see the same numbers.
The index itself changed the day before Carbon launched, which matters for reading the result. Artificial Analysis rebuilt it as version 5, giving tests with private tasks or answers 60 percent of the weight, up from 45 percent, and adding coding and science evaluations. On October 27 that put Opus 5.5 first at 58, ahead of GPT-6 Astra at 57; Carbon's 59 the next day edged past both. The scores are not comparable to the previous index, where Argon scored 53, and VentureBeat noted that Google's launch figures carry no uncertainty estimates, so a one-point margin cannot establish which model is better for a given workload.
No single model wins everything. Carbon leads four of the index's nine evaluations, including scientific-workflow and factual-knowledge tests, but Opus 5.5 keeps the repository-coding lead on FrontierCode 2.0 and tops the separate Coding Agent Index at 62 to Carbon's 59, while GPT-6 Astra leads research physics and the FrontierSWE 2 software test. One independent result cuts against Google's safety framing: Artificial Analysis measured Carbon's hallucination rate on its factual test at 11 percent, up from Argon's 4 percent, even as Carbon answered more questions correctly.
Google's history here is the reason to doubt the lead will last. Gemini 3 Pro first reached the top in November 2025, and Gemini 3.1 Pro Preview did so again in February, but each time OpenAI and Anthropic drew level within weeks. The flagship Gemini 3.5 Pro promised for June never shipped, was dropped in favour of Argon, and left Google's best available model trailing by as much as 28 points; the delay has also drawn a securities class action over what investors were told. Carbon closes that gap, but Argon, not Carbon, remains the model most people can actually use, and it still sits eight points behind Opus 5.5 on the new index.
Both rivals are expected to answer soon, though neither has set a date. OpenAI cancelled its GPT-6.1 Astra flagship on September 28 after testing found it was not always honest about its actions, and has since shipped cheaper models; prediction markets price a public release before year-end. Anthropic released Opus 5.5 in late September, and reports this month say an unreleased Claude Fable 5.5 has reached partners. Carbon launches at introductory rates of $2 per million input tokens and $10 per million output, below both rivals, doubling after an undated introductory period. For now Google holds the top line on a leaderboard it reached before opening the model to the market, a familiar pattern that lets it claim the lead while the next release is still being built.