Google has announced Gemini 4 Carbon, a new frontier AI model positioned as an improvement on Gemini 4 Argon. The announcement highlights coding and cybersecurity results, but most people cannot use Carbon yet: Google says it is taking a deliberative approach to testing before a wider release.
A step beyond Argon, with a narrow gap in tests
Carbon follows Gemini 4 Argon by two weeks. It is not the Flash model equivalent that some users have been waiting for. Instead, Google presents it as a stronger model in the same general class as Argon.
The company’s benchmark results put Carbon slightly ahead of Argon across the tests it shared. Google also reports that Carbon beats GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 in almost every test. The differences are only a few points, however, and all the frontier models land in a similar range.
That makes the results useful as a snapshot of how Google compares these systems, but not a complete account of how they perform in everyday use. Benchmarks measure performance on selected evaluations; they cannot by themselves settle how useful a model will be across the many tasks people might give it.
Google describes Carbon as a new model, though the available information does not establish how it was built. It could be a post-training upgrade of Argon, but that remains uncertain. What Google does say is that Carbon has improved on Argon by slim margins in every test it has provided.
Cybersecurity is central to the pitch
Google points to cybersecurity as one of Carbon’s most notable areas of progress. It says the model makes a “significant leap” in vulnerability detection, validation and remediation. The company also says Carbon is twice as likely as Argon to generate code free of vulnerabilities.
The Chrome team claims Carbon is 10 times better at generating “high-quality” patches compared to non-Google models. Google also reports that Carbon’s score on WebDev Arena has increased by 89 points, putting it at number four on the leaderboard.
These claims frame Carbon as a tool that could help find and repair software weaknesses. They also help explain why Google is treating access cautiously: capabilities for agentic coding and cybersecurity can have important consequences, and the company says it wants to evaluate the model before making it broadly available.
Google says Carbon resisted prompt-injection attempts
Google presents safety results alongside the capability claims. In the Grey Swan evaluation, Carbon received a 0 percent score for attempted prompt-injection attacks, which Google describes as resisting every attempt. The reported figures were 0.7 percent for Argon, 1 percent for Opus 5.5 and 8.5 percent for GPT-6 Astra.
Those numbers offer one point of comparison across the listed models. They are results from a specific evaluation, rather than a guarantee about every possible interaction. Google’s broader message is that Carbon’s added capabilities come with safeguards it considers more rigorous.
Testing comes before wider access
Google says the release plan will resemble Argon’s staged rollout. The company is working within a voluntary testing framework established by the US government, and plans to distribute Carbon through its Fairwinds cyber defense program after that process is complete.
There is no announced date for general availability. If Carbon follows the same pattern as Argon, it could arrive in a couple of weeks, but Google has made no promise that it will. The limited initial audience is expected to include AI Ultra subscribers and developers using the Gemini API.
Google says the introductory API rates will be $2 per million input tokens and $10 per million output tokens, after which each price will double. The company has not specified when Carbon’s introductory pricing will end. Argon’s discounted rates are only good through the end of November, but Google has declined to provide an end date for Carbon, suggesting it has not decided when the increase will take effect.
For now, the announcement sets expectations more than it opens access. Google is presenting Carbon as a small but consistent step beyond Argon, with particular emphasis on cybersecurity, while keeping the model within a testing and staged-release process. Users and developers will have to wait for the company to complete that process and clarify the timeline.