Gemini 4 Carbon, which was released two weeks ago but was only available to cybersecurity defense organizations, was finally made available to paying developers and enterprise customers on Tuesday. The entrance is the Gemini API in Google AI Studio. Enterprises use Gemini for business, and individual users who subscribe to AI Ultra can also choose it in the Gemini application.

Price is the most important thing to keep an eye on this time. The initial price is US$2 per million input tokens and US$10 per million output tokens, which is exactly the same as Gemini 4 Argon in the same category. But this price only lasts until November 30, and will directly become $4 and $20 starting in December. In other words, the cost calculated by running it for evaluation now will be multiplied by two next month.

In terms of specifications, the context has 1.04 million tokens, and the regular output limit is 65,000. Turning on Long Decode can also increase the output to 1.04 million; the inference intensity is divided into three levels. The division of labor between it and Argon is: Argon is in charge of daily engineering, while Carbon is willing to burn more reasoning tokens to test hypotheses and do verification, leaving slow work such as multi-step debugging and open scientific research.

The scores given by Google focus on security. SEC-bench Pro vulnerability discovery was 88.2%, CWE-bench remediation was 71.0%, and the attack success rate on Gray Swan's prompt injection benchmark was 0%. Wiz's security team said it took less than two hours to find a zero-day vulnerability. On the programming side, CursorBench 4.0 ranks first.

Hassabis himself gave a more understandable example. When he was 17, he co-wrote the 1994 game "Theme Park". This time he threw the compiled game to Carbon. The model figured out the operating logic by itself, added a DeepMind-themed ride, and then ported it to the web page at once. Last year he did something similar with Gemini 3 and stayed up late tinkering with it for hours.

Antigravity also previews /swarm multi-agent collaboration simultaneously.

For those who want to try it, these twenty days are the most cost-effective window. But when you budget, don’t think in terms of $2, think in terms of $4.