Google releases Gemini 4 Carbon, strengthening long-duration development and vulnerability discovery and remediation capabilities

On October 28, 2026, Google announced a new AI model, "Gemini 4 Carbon." It says the model has enhanced capabilities in long-duration development work spanning multiple processes, as well as in cyber defense from vulnerability discovery to remediation. In the week of the announcement, it began offering the model to some cyber defenders and will gradually expand the scope. It indicated a policy of eventually providing it to developers, enterprises, and general users as well.

Gemini 4 Carbon: our next frontier model - Google

Continues development work for several days⁠, also improves reliability of answers and execution

Gemini 4 Carbon is a model built on Gemini 4 Argon[1], which became available on October 13, with enhanced coding ability and ability to advance work while using tools. Within Google, there were cases where Carbon spent several days investigating and fixing the causes of bugs lurking in large codebases and low-level systems. It reportedly kept working even when it got stuck partway through.

Google says it has also enhanced the ability to act within the scope of a request, reduce cases of not sufficiently performing necessary work, and carry work through to the end while respecting project constraints. When instructions are ambiguous, it reportedly investigates related materials and, if there are discrepancies, informs the user before execution.

In internal tests, unsupported claims were reduced to about half compared with Argon. If there are limits to the methods of investigation or verification, it states them in the deliverables, and for processes whose execution conditions are not met, it indicates what is missing and refrains from proceeding. It also says that on Terminal-Bench 4.0, while increasing the correct-answer rate on tasks, it reduced the rate of successful cheating in evaluations to 0.0% and the rate of attempted cheating to 0.4%.

According to Google, on the Artificial Analysis Intelligence Index v5.0, which comprehensively evaluates AI model capabilities, Carbon recorded a score of 59 and took first place. In the comparison table the company presented at the time of announcement, Argon's score was 50.

Gemini 4 Carbon is our next frontier model.

It’s state-of-the-art across cybersecurity and long-horizon capabilities.— Google DeepMind (@GoogleDeepMind) October 28, 2026

Google's model evaluation materials include, in addition to this overall evaluation, results for Terminal-Bench-family benchmarks measured by Artificial Analysis. For the coding-field Terminal-Bench 4.0 (AA) and the science-field Terminal-Bench-Science 0.1 (AA), comparisons with other companies' models are as follows.

Model Terminal-Bench 4.0 (AA) Terminal-Bench-Science 0.1 (AA)

Performance evaluated in vulnerability discovery⁠, exploitability verification⁠, and remediation

Google says Carbon's capabilities improved at each stage of cyber defense, from vulnerability discovery to exploitability verification and root-cause remediation. Exploitability verification is conducted so defenders can determine whether a vulnerability needs to be fixed and which ones to prioritize.

In internal benchmarks, the success rate of generating secure code reportedly exceeded twice that of Argon. In the following three benchmarks, performance was evaluated under conditions with the model's safety guardrails removed. SEC-bench Pro and ExploitGym were evaluated by Google, and CWE-bench v1 by Collinear AI.

In vulnerability discovery (SEC-bench Pro), the model tackles tasks to find vulnerabilities in a JavaScript engine and create verification code that reproduces them. The success rate in a single attempt was 88.2% for Carbon and 83.4% for Argon[2].

In exploitability verification (ExploitGym), it evaluates whether the model can create code that exploits a known vulnerability. When the time limit per task was set to 10 hours, the success rate was 54.8% for Carbon and 42.0% for Argon. Under the standard 6 hours, they were 53.5% and 39.5%, respectively. During evaluation, the model was not connected to the internet, and mitigations that hinder attacks in the evaluation environment were disabled.

In vulnerability remediation (CWE-bench v1), the model tackles tasks to audit code and fix vulnerabilities. The success rate in a single attempt was 71.0% for Carbon and 68.0% for Argon[3].

For detailed evaluation conditions and judging methods, refer to the aforementioned model evaluation materials.

Resistance to prompt injection also improved

Google says resistance to indirect prompt injection also improved. In this attack, malicious instructions are embedded in documents handled by the model, attempting to make it deviate from the original request. On Gray Swan's IPI benchmark, under conditions allowing 15 attack attempts, the attack success rate dropped from 0.7% for Argon to 0.0% for Carbon.

Staged rollout to defenders⁠, expanding to developers and general users as well

Google is providing Gemini 4 Carbon first to some cyber defenders and will sequentially expand the scope of provision to trusted partners through the Fairwind Program. The program is an initiative to provide models to defenders first, given that capabilities useful for discovering and fixing vulnerabilities could also be misused for attacks. It targets governments, medical institutions, telecommunications carriers, and others, and participation requires screening.

Going forward, after continuing to collect feedback and improve guardrails, the company aims to provide the model to developers, enterprises, and general users through its various products as soon as possible[4].

The announced introductory pricing is $2 per million tokens for input and $10 for output, with cached input to be discounted by 95% off the input price. After the introductory period ends, it will be $4 for input and $20 for output.