Gemini 4 Carbon ranks first on Artificial Analysis; access is limited

Carbon builds on Gemini 4 Argon, which Google launched two weeks earlier, and Google says it illustrates the pace of its frontier work: researchers and models advancing capabilities together while safeguards are strengthened in tandem. The headline number is the Artificial Analysis Intelligence Index v5.0, where Artificial Analysis's own leaderboard lists Carbon (high reasoning) at 59, ahead of Claude Opus 5.5 (max) at 58, GPT-6 Astra (max) at 57, and Argon (high) at 50. Artificial Analysis is an independent evaluator, and its page notes Carbon has no publicly available API yet. VentureBeat, citing Artificial Analysis, adds that Carbon's hallucination rate on the AA-Omniscience test rose from 4% to 11% versus Argon even as its overall score there led.

The work it is built for is long-horizon engineering: Google says internal teams ran Carbon for days at a time on complex, multi-step tasks such as root-causing and patching subtle bugs across large codebases and low-level systems. Google says web development improved too, with Carbon gaining 89 Elo points to reach fourth place (1768) on WebDev Arena. On internal benchmarks, Google reports more than double Argon's success rate at generating secure code.

Google's benchmark table compares Carbon against Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 across 17 rows. As reproduced by VentureBeat, Carbon posts the top score in 15 of them. On FrontierSWE 2, GPT-6 Astra scores 65.5% and Claude Opus 5.5 62.3%, against Carbon's 61.2%. On the OSWorld computer-use test, scored on an 82-task offline subset, Claude Opus 5.5 reaches 87.2% to Carbon's 81.5%. Google publishes its methodology on DeepMind's evals page and says non-Gemini results come from providers' self-reported numbers unless noted.

In the other rows, per VentureBeat's reproduction of the table, Carbon scores 70.4% on the Vals Index, 65.2% on Terminal-Bench 4.0, 77.4% on Agents' Last Exam, and 96.2% on GraphWalks, which Google reports for contexts of 256,000 to 1 million tokens and computed itself. Some margins are narrow: on the Vals Index Carbon's 70.4% is 0.2 points above Claude Opus 5.5's 70.2%.

The three cybersecurity figures are the ones Google leads with, and they carry a condition that has to travel with them: Google says they were evaluated without guardrails. SEC-bench Pro: 88.2% for finding real vulnerabilities. ExploitGym: 54.8% for exploitation and validation. CWE-bench: 71.0% for patching at the root cause. Google computed SEC-bench Pro and ExploitGym itself, and CWE-bench was evaluated by Collinear AI. These figures describe the model with safeguards off; Google's announcement does not say what the guarded product will score.

Two more claims need the same handling. Google says its Chrome team reported Carbon performing more than 10 times better than another lab's large frontier model at generating high-quality patches; Google does not name the model or publish the underlying results. And the safety figures are Google's own internal counts: high-severity misalignment alerts down almost 80% in live coding, unsupported claims cut by about half versus Argon, and a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark, down from Argon's 0.7%.

The announcement's own summary comes from Koray Kavukcuoglu, SVP at Google DeepMind and Google's Chief AI Architect, who calls Carbon “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” Google says that starting this week it will make Carbon available to a subset of cyber defenders and gradually phase in additional trusted partners through its Fairwind Program. Google also says it is engaged in the U.S. government's voluntary process for pre-release model testing.

Google says it aims to make Carbon available to developers, businesses, and consumers through its product surfaces “as soon as possible,” after gathering feedback and iterating on guardrails, and gives no date. It lists introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper, the same as Argon's introductory rates. The higher rates of $4 per million input and $20 per million output apply after the introductory period expires; Google has not stated when that is, and the announcement describes no usage threshold. VentureBeat notes that Argon's introductory period ends November 30.