Google has announced Gemini 4 Carbon, a new AI model it says improves on its recently released Gemini 4 Argon in coding, long-running engineering work and cybersecurity. For now, though, the Gemini 4 Carbon cybersecurity pitch is mostly for a limited group of cyber defenders, not an immediate public launch.
Google said on October 28 that it will initially give access to a subset of cyber defenders, then gradually extend it to trusted partners in its Fairwind Program. The company said it is still collecting feedback and adjusting safeguards before offering Carbon to developers, businesses and consumers. It has not provided a broad-release date.
Carbon arrives two weeks after Argon. Google describes the new model as designed for sustained agent-style work, meaning software that can carry out a sequence of tasks rather than answer one prompt and stop. Its internal teams, Google said, have watched Carbon work for days while tracing and patching subtle defects across large codebases and low-level systems.
What do Gemini 4 Carbon’s cybersecurity benchmarks measure?
Google reported results across three security tasks. On SEC-bench Pro, which tests whether a model can identify vulnerabilities in real software, Carbon scored 88.2%. It scored 54.8% on ExploitGym, a test of whether a known flaw can be turned into an exploit. On CWE-bench, which evaluates fixes for software weaknesses, Google reported a 71.0% result.
The company said those external benchmark evaluations were conducted without its guardrails. That detail matters: the tests measure technical capability in defined tasks, not whether the system should carry out a request in a deployed product.
Google also said Carbon had more than twice Argon’s success rate in its internal tests for producing secure code. It said Chrome engineers found it generated high-quality patches more than 10 times as effectively as a frontier model from another lab. Those are Google’s internal findings, not independently audited comparisons.
Ars Technica noted that benchmark tables do not settle real-world performance and that the reported separation between leading models in the discussed tests can be only a few points. Google’s claims therefore show where it believes Carbon performs well, rather than a general proof that it will be better for every security team or production environment.
When will Gemini 4 Carbon be available?
The model is available first to selected cyber defenders, according to Google. Wider access through the Fairwind Program is planned in stages, while developers, companies and consumers remain waiting. Google said it is participating in the U.S. government’s voluntary pre-release model-testing process as it handles the rollout.
Available now: a subset of cyber defenders.
Next: additional trusted partners through Fairwind.
Not yet generally available: developers, businesses and consumers.
Google set introductory API prices at $2 per million input tokens and $10 per million output tokens. Cached input tokens will cost 95% less than the regular input-token rate, the company said. Pricing explains the meter, not the access: most prospective customers cannot yet use the model.
This story draws on original reporting from The Verge.
