Google’s Gemini 4 Carbon Targets Cybersecurity’s Hardest Problems

Google has announced Gemini 4 Carbon, a new frontier model aimed at cybersecurity, complex software engineering, and tasks that can run for long periods. The model arrives two weeks after Gemini 4 Argon, making Carbon the newer system despite Argon being only three weeks old in some comparisons.

Google calls Carbon “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” The company says it marks a “significant leap” in finding, validating, and fixing vulnerabilities, while also improving coding, computer use, knowledge work, science, and math.

A careful rollout for cyber defenders

Carbon will not open to everyone at once. Google plans to give access to a subset of cybersecurity defenders through its Fairwind Program, then expand access to additional trusted partners. The company is working within the U.S. government’s voluntary framework for pre-release model testing and says it wants to gather feedback and iterate on guardrails before a wider launch.

Google has promised access for developers, businesses, and consumers “as soon as possible,” but has not provided a date for broad availability. The first wider group will probably include AI Ultra subscribers and developers using the Gemini API, although that plan is not confirmed as a public release schedule.

The approach mirrors the release plan for Argon. Google says its teams have seen Carbon “run for days,” while Varun Mohan wrote that people have enjoyed using the model internally on Antigravity to solve their hardest tasks.

Carbon supports text, image, video, and speech inputs, with text output. It has a 1M-token context window and a 1M maximum output. The announcement does not provide an equivalent API specification for features such as function calling, code execution, structured responses, or computer interaction.

Cybersecurity results lead Google’s pitch

Google’s evaluations show Carbon scoring 88.2% on SEC-bench Pro for vulnerability finding, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Google also says Carbon is twice as likely as Argon to generate secure code and that its Chrome team found Carbon produced “high-quality” patches 10 times better than non-Google models.

Safety results focus on reducing shortcuts and loophole exploitation. Successful cheating on Terminal-Bench 4.0 fell to 0.0%, while cheating attempts fell to 0.4%. Unsupported claims dropped about 2 times compared with Argon, overreach fell 26%, and under-service fell 29%.

Google also says Carbon resisted every attempted prompt-injection attack in the Grey Swan evaluation. Its 0 percent result compares with 0.7 percent for Argon, 1 percent for Claude Opus 5.5, and 8.5 percent for GPT-6 Astra.

There is a tradeoff. Carbon’s hallucination rate on the Omniscience evaluation rose from Argon’s 4% to 11%, though that remained the second-lowest measured result. Claude Opus 5.5 and GPT-6 Astra both scored 50%. That higher hallucination rate has raised concerns among users about Carbon’s use in cybersecurity.

Strong benchmark showing, with clear limits

Artificial Analysis places Carbon first on Intelligence Index v5 with a score of 59. Carbon leads four of the nine index evaluations: HLE-Diamond, TB-Science, AA-Omniscience, and AutomationBench. It also leads Omniscience with 54 and ranks first on the Vals Finance Index and Legal Research Benchmark, while placing second on the Vals Index.

On the Coding Agent Index, Carbon scores 59 through Antigravity CLI, compared with 62 for Claude Code with Opus 5.5 [max]. Carbon scores 51% on FrontierCode v1.1, behind Opus at 56%. On AA-Briefcase, it reaches 1609 Elo, an increase of 126 points from Argon, but still trails Grok 4.7, Muse Wonder, and Kimi K3.1 Pro at 1678.

The coding-agent results were measured in Antigravity CLI, so results in other harnesses remain untested. Carbon scores 61.2% on FrontierSWE, behind GPT-6 Astra at 65.5, Opus 5.5 at 62.3, and Sonnet 5.5 at 61.9. It does set a narrow state of the art on SWE-sweep at 9.6%.

Carbon also ranks first on Vending-Bench 2 from Andon Labs and first on Blueprint-Bench 2, where it ties the human baseline for 3D spatial understanding. Arena places Carbon first in Text Arena with 1528, second in Agent Arena at a $1.80 median cost per task, and fourth in Code Arena: WebDev with 1768. Arena attributes the gains over Argon to more praise, more confirmed successes, and better steerability.

Carbon costs $7.76 per index task, compared with $10.04 for Opus 5.5 and $6.25 for GPT-6 Astra. It uses 123k output tokens per task, 29% fewer than Opus, but produces 64% more output tokens than Argon, with most of that increase appearing on finance, tax, and legal tasks.

Pricing starts low, but the end date is unclear

Google will launch Carbon at $2 per million input tokens and $10 per million output tokens. Cached inputs receive a 95% discount, bringing the cost to $0.20 per million tokens. The standard prices are listed as $4 per million input tokens and $20 per million output tokens, meaning the introductory rate will double when that period ends.

Argon’s discounted rate runs through November 30, 2026, but Google has not given Carbon an end date. The company has declined to provide one, suggesting the timing of the price increase has not been decided.

Google says Carbon shows the pace at which it is pushing the frontier forward. Still, scaling01 estimates that Google remains about two months behind the next releases from OpenAI and Anthropic. For now, Carbon’s mix of cybersecurity gains, long-horizon operation, and restricted access makes it a model to watch before it becomes a model everyone can use.

Stay connected via Google News

Follow us for the latest AI News