Gemini 4 Carbon shines on security tests, but a broad rollout is still some way off
Two weeks after Argon, Google is already presenting a stronger successor, which will first go only to selected cyber defenders.
Google on Wednesday announced Gemini 4 Carbon, barely two weeks after launching Gemini 4 Argon. According to Ars Technica, the model makes its biggest gains in cybersecurity: it is said to find and fix vulnerabilities better and to write code free of flaws twice as often as Argon. Google's Chrome team claims Carbon delivers patches ten times better than a large model from another AI maker. Ars notes that the gap with the competition amounts to just a few points in most tests, and that it is unclear whether this is a genuinely new model or an updated Argon.
Ordinary users will have to wait. Google first wants the model tested under the US government's voluntary framework and then distributed through a cyber defence programme, Ars Technica writes. The Register reports that an unspecified group of security specialists will get access this week. For developers, Carbon costs 2 dollars per million input tokens and 10 dollars per million output tokens during an introductory period with no fixed end date; after that the price doubles.
That eventual price matches Anthropic's Claude Opus 5.5, which Google clearly sees as its main rival, according to The Register. In the Artificial Analysis rankings Carbon tops the overall index and is second on the coding agent index, behind Opus 5.5. On the Terminal-Bench 4.0 coding test Carbon scores 65 percent, according to Heise. Google chief executive Sundar Pichai said alongside the quarterly results that the company's own models now process nearly 32 billion tokens per minute, more than four times as many as a year ago. Transparency: this edition was compiled by Claude, made by Anthropic.