Google has announced Gemini 4 Carbon, a new artificial intelligence model designed for complex software engineering and cybersecurity tasks, reporting improved performance in identifying vulnerabilities and producing software fixes.
The company introduced the model on October 28.
Gemini 4 Carbon follows the release of Gemini 4 Argon earlier this month.
Google says the new model can work on extended engineering tasks, including investigating difficult software problems and producing patches across large codebases.
Initial access will be restricted to selected cybersecurity defenders and trusted partners.
Cybersecurity benchmarks show different results
Google published performance figures for three cybersecurity benchmarks.
Gemini 4 Carbon achieved 88.2% on SEC-bench Pro, a benchmark used to assess vulnerability detection.
The company reported a score of 54.8% on ExploitGym, which examines the ability to validate whether vulnerabilities can be exploited.
Carbon also achieved 71.0% on CWE-bench, which measures software patching performance.
These benchmarks test different capabilities.
The percentages cannot be averaged to produce a meaningful overall cybersecurity score.
| Benchmark | Reported score |
|---|---|
| SEC-bench Pro | 88.2% |
| ExploitGym | 54.8% |
| CWE-bench | 71.0% |
Google said the external cybersecurity benchmark results were obtained without deployment guardrails.
That distinction matters when comparing research performance with the behaviour of a model made available through a commercial product.
Model designed for longer engineering tasks
Google says Gemini 4 Carbon can work through complicated software problems over extended periods.
Internal tests involved tasks such as identifying the causes of software failures and developing fixes across large codebases.
The company reported that the model performed better than Gemini 4 Argon in several internal evaluations.
On WebDev Arena, Google reported a rating of 1,768 and a fourth-place position.
The company also said Carbon had improved its ability to generate secure code.
Some of those results come from Google’s own internal testing and have not been independently reproduced in the announcement.
Google restricts early access
The company is initially making Gemini 4 Carbon available to a selected group of cybersecurity defenders.
Further access will be provided to trusted partners through Google’s Fairwind programme.
Google said it is participating in a voluntary United States government process for testing advanced models before release.
The company plans to make Carbon available more widely to developers, businesses and consumers after additional testing.
No confirmed general release date was announced.
Introductory pricing will eventually double
Google disclosed introductory API pricing of US$2 per million input tokens and US$10 per million output tokens.
The company also announced a 95% discount on cached input tokens.
At the introductory rates, processing one million uncached input tokens and generating one million output tokens would cost US$12.
That calculation excludes any additional charges associated with other services or tools.
Google’s pricing note states that the standard rates after the introductory period will be US$4 per million input tokens and US$20 per million output tokens.
The equivalent uncached input and output workload would therefore cost US$24 at those later rates.
Introductory input: US$2 per million tokens.
Introductory output: US$10 per million tokens.
Later input price: US$4 per million tokens.
Later output price: US$20 per million tokens.
Introductory period end date: Not specified in the announcement.
Safety evaluations remain limited to tested conditions
Google also published results from evaluations intended to measure model behaviour under difficult instructions and attempted misuse.
The company reported a 0% successful attack rate on a particular indirect prompt-injection benchmark.
That result means no successful attacks were recorded under the conditions of that evaluation.
It does not establish that the model is immune to every possible prompt-injection technique.
Google also reported reductions in unsupported claims and other undesirable behaviour during internal testing.
The company said it would continue evaluating the model and adjusting its safeguards before expanding access.
The October 28 announcement provides benchmark results and proposed pricing, but the model has not yet entered unrestricted public availability.
Google, Introducing Gemini 4 Carbon, 28 October 2026
