Google’s Gemini 4 Carbon tops AI rankings, but only cyber defenders get it

Google has released Gemini 4 Carbon, its newest frontier AI model, to a small group of cyber defenders. It arrives two weeks after Gemini 4 Argon reached paying users. Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief AI architect, announced it in a Google blog post on Wednesday.

Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. The Artificial Analysis index puts Anthropic’s Claude Opus 5.5 second with 58, and OpenAI’s GPT-6 Astra third with 57.

“Carbon illustrates the pace at which we’re pushing the frontier forward,” Kavukcuoglu wrote.

Google says Carbon leads three outside cybersecurity tests, which it ran without guardrails. It scored 88.2% on SEC-bench Pro, which measures finding real vulnerabilities. It scored 54.8% on ExploitGym, which tests whether a flaw can be exploited. On CWE-bench, which measures patching, it scored 71.0%.

On Google’s internal tests, Carbon was more than twice as successful as Argon at writing secure code. Google’s Chrome team found it more than 10 times better than another lab’s large frontier model at writing high-quality patches. Google did not name the lab.

Google says its teams have seen Carbon run for days on long engineering tasks. These include tracking down and patching subtle bugs across large codebases. Carbon also climbed 89 points to fourth place on WebDev Arena, a leaderboard for AI-built websites.

Google reports that Carbon makes about half as many unsupported claims as Argon. In its internal coding environments, high-severity misalignment alerts fell by almost 80%. Misalignment means the model acting in ways its developers did not intend.

On Gray Swan’s prompt injection benchmark, no evaluated attack succeeded against Carbon. Argon’s attack success rate was 0.7%. On Terminal-Bench 4.0, Carbon’s rate of successful cheating was 0.0%.

Google’s own comparison table shows two tests where Carbon does not come first. GPT-6 Astra leads on FrontierSWE 2, a software engineering test, with 65.5% against Carbon’s 61.2%. Claude Opus 5.5 leads on OSWorld, which measures computer use, with 87.2% against 81.5%.

Artificial Analysis ranks Carbon second on its Coding Agent Index, behind Claude Code running Opus 5.5. On its knowledge test, Carbon answered 58% of questions correctly, up from 48% for Argon, Carl Franzen reported for VentureBeat . Its hallucination rate on the same test rose from 4% to 11%.

That rate counts wrong answers among those Carbon did not get fully right. Google has not published a technical report for Carbon yet, VentureBeat noted.

Starting this week, Carbon goes to a subset of cyber defenders. Google will add more trusted partners through its Fairwind Program. It is also taking part in the US government’s voluntary process for testing models before release. Developers, businesses and consumers will get Carbon “as soon as possible”, Google says.

Argon took the same path. Google gave cyber defenders Argon first on 30 September, then opened it to everyone who pays on 13 October.

Carbon will launch at Argon’s introductory price of $2 per million input tokens and $10 per million output tokens. A footnote in Google’s post says prices will later rise to $4 and $20. Google declined to say when, Ryan Whitwam reported for Ars Technica .

Google announced Carbon on the day Alphabet reported its third-quarter results , a day after DeepMind released the Genie 4 world model . Sundar Pichai, chief executive of Google and Alphabet, brought up Carbon in his earnings call remarks .

“With Gemini 4 Argon and now Carbon, which we announced today, we’re excited to be leading at the frontier,” Pichai said.