Google introduced Gemini 4 Carbon on October 28, 2026, focusing the model on coding, long-running agentic tasks and cybersecurity. The company plans to make it available first to a subset of cyber defenders, followed by developers, businesses and consumers after it gathers feedback and adjusts safeguards.
Coding and long-running tasks
In internal use, Carbon worked for days on multistep engineering tasks, according to Google. Its teams used it to find the causes of subtle bugs and patch them across large codebases and low-level systems.
In internal tests of secure code generation, the model achieved more than twice the success rate of Gemini 4 Argon, which Google introduced two weeks earlier, the company said. Google also said the Chrome team reported that Carbon performed more than 10 times better at generating high-quality patches than a large advanced model from another lab, which it did not name.
Google says Carbon reached fourth place in WebDev Arena with 1,769 points, following an increase of 89 Elo points. The company links the result to improved output quality in web development. According to Google, the model also ranked first with a score of 59 in the updated version 5.0 of the Artificial Analysis Intelligence Index.
Finding and fixing vulnerabilities
Google also published results from external cybersecurity tests in which Carbon was evaluated without safeguards. The company said the model scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for checking whether vulnerabilities can be exploited, and 71.0% on CWE-bench for fixes.
Reliability and resilience in tests
In internal testing, Carbon produced approximately half as many unsupported claims as Argon, according to Google. In live coding environments with the company’s internal teams, alerts for high-severity misalignment fell by almost 80%.
Google says the rate of successful cheating in Terminal-Bench 4.0 fell to 0.0%, while attempted cheating fell to 0.4% and overall accuracy improved. The company also reports a 26% reduction in overstepping the task and a 29% reduction in failing to meet requirements.
In Gray Swan’s test of resistance to indirect prompt injection, the attack success rate against Carbon was 0.0%, according to Google. The corresponding rate for Argon was 0.7%.
Gradual access to the model
Google will begin making the model available to a subset of cyber defenders during the week of the announcement, followed gradually by additional trusted partners in its Fairwind Program. The company plans access for developers, businesses and consumers after gathering feedback and adjusting safeguards.
Carbon’s planned introductory rates match Argon’s: each million input tokens will cost $2, while a million output tokens will cost $10. Cached input tokens will cost 95% less than regular input tokens.
