Google announced Gemini 4 Carbon on Wednesday, barely two weeks after it shipped Gemini 4 Argon. The company is calling Carbon its newest frontier model, and it's leaning hard on gains in three areas: coding, agentic tasks, and cybersecurity defense, specifically finding vulnerabilities in code.
Google says Carbon now sits at the top of the recently refreshed Artificial Analysis Intelligence Index, version 5.0, with a score of 59. That's an independent benchmark used across the industry to compare how AI models reason, code and follow instructions, and claiming the top spot there is the kind of number Google wanted to lead with.
The company attributes the speed of this release to what it describes as researchers and models "working hand in hand to advance capabilities faster." In plain terms, Google is using its own AI tools to help build the next AI tool, and it's happy to say so out loud.
What actually changed
Google says it took feedback from the Argon launch and pushed Carbon further on coding, agentic workflows and scientific tasks, along with full-stack web development and multimodal understanding, meaning how well the model reads and reasons across text, images and other formats at once.
On safety, Google says it worked on alignment, the ongoing effort to keep a model's behavior matching what its developers actually intend, and trained Carbon to "respond to complex or high-pressure tasks in the right way." The company hasn't published specifics on what that training involved or how it's measured.
Rollout and pricing
Carbon isn't broadly available yet. Google says it's "coming soon to developers, businesses, and consumers," with no firm date attached to that promise. The first group to get access will be "cyber defenders," starting this week, a nod to the model's advertised strength in catching security vulnerabilities before attackers do.
When it does roll out more widely, Google is setting introductory pricing at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced 95% below the standard input rate. Tokens are the chunks of text a model processes, and pricing by the million is standard practice for API access aimed at developers rather than everyday app users. Once the introductory window ends, the rate climbs to $4 per million input tokens and $20 per million output tokens, doubling the entry price.
Google hasn't said how long the introductory pricing will last, or when consumer-facing products like the Gemini app will actually get Carbon under the hood. Argon's own rollout gives some indication of pace, since Google went from that launch to this one in just fourteen days. If that cadence holds, Carbon's run as the flagship model might not last much longer than its predecessor's did.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.