Google Unveils Gemini 4 Carbon, an AI Model Few Can Use Yet

Google announced Gemini 4 Carbon on Oct. 28, a new frontier AI model the company says outperforms its own two-week-old Gemini 4 Argon, along with rival systems GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 in most benchmark tests, according to Ars Technica. Most people cannot try it yet.

What is Gemini 4 Carbon?

Carbon is Google's latest large language model, arriving just two weeks after Argon's rollout, Ars Technica reported. Google frames Carbon as a new frontier model rather than the lighter-weight "Flash" variant users have been anticipating. The outlet notes the two models "have a lot in common," with Carbon beating Argon "by slim margins in every test Google has given it," raising the possibility that Carbon is a post-training upgrade rather than a ground-up build.

How does Carbon compare to Argon on security and coding?

Google says Carbon shows a "significant leap" in vulnerability detection, validation and remediation, and is twice as likely as Argon to generate code free of vulnerabilities, per the Ars Technica report. The Chrome team claims Carbon is 10 times better than non-Google models at producing "high-quality" security patches. Carbon also gained 89 points on the WebDev Arena leaderboard, moving it to fourth place overall, according to the report.

Where does Carbon rank against GPT-6 Astra and Claude?

Google's internal benchmarks place Carbon ahead of GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 in nearly every test cited, Ars Technica reported. But the outlet cautions that "every frontier model scores in the same neighborhood," with gaps between Carbon and competitors amounting to only a few points. On the Grey Swan prompt-injection evaluation, Carbon resisted every attack for a 0 percent failure rate, compared with 0.7 percent for Argon, 1 percent for Claude Opus 5.5 and 8.5 percent for GPT-6 Astra, according to the report.

When can developers and consumers use it?

Google has not set a firm release date, saying Carbon's capabilities in agentic coding and cybersecurity require a "deliberative approach," Ars Technica reported. The company said it is working inside the voluntary testing framework set up by the U.S. government and plans to deploy Carbon through its Fairwinds cyber defense program once testing wraps. If the rollout follows Argon's pattern, broader availability could land within a couple of weeks, with AI Ultra subscribers and developers using the Gemini API getting access first, according to the report.

What will Carbon cost to use?

Google plans a temporary discounted API rate of $2 per million input tokens and $10 per million output tokens, matching the introductory pricing it used for Argon, Ars Technica reported. That rate is set to double once the discount period ends, though Google has not specified when that will happen.

The staggered release mirrors a broader industry pattern of AI firms expanding paid access tiers around new models, a dynamic also visible in Meta's recent AI-focused subscription plans.