Google announced in the early morning of the 11th Taiwan time that its most powerful model, Gemini 4 Carbon, is open to all paid API developers and enterprises. Google AI Ultra subscribers can also use it in the Gemini App and search AI mode.

When Carbon was released on October 28, it was only open to a limited number of security defense personnel. Chain News at the timereportI have written about its evaluation results and security capabilities. This time Google has added the open scope, and the API price page also states the expiration date of the preferential price.

The API discount is only available until November 30th and will double from December 1st

according to Google AI Studio pricing page, Carbon has no free quota. The payment plan charges US$2 for every 1 million input tokens and US$10 for output tokens (including tokens used for model thinking). This price is only available until November 30; starting from December 1, it will be adjusted to US$4 for input and US$20 for output, and the cache input will also be adjusted from US$0.1 to US$0.2. The previous generation Argon price page stated the same set of schedules.

Although the unit price is the same as Argon, Google isOfficial blogIt is said that Carbon will use more tokens than Argon for exploration, reasoning and repeated verification. Google’s recommendation is that Argon is enough for most software engineering and knowledge work, and the most difficult tasks such as multi-step program debugging and open scientific research should be left to Carbon.

Most users want the Ultra plan, but the blog doesn’t mention other plans.

On the consumer side, Google writes that it will be pushed to "all Google AI Ultra subscribers", including Gemini App, search AI model and development tool Antigravity. The article does not mention other subscription plans. Which models can be used in each plan can be compared Complete Guide to Google Gemini。

Enterprises can use it through Gemini for business API or model menu. Four weeks ago, Google opened an application for 10,000 scientists to use Argon for free for six months. Now selected candidates can also use Carbon. Applications are still open and reviewed in batches every week.

Google says CursorBench 4.0 ranks first, costing less than half of its competitors per task

Google says Carbon ranks first in Cursor's CursorBench 4.0, Perplexity's research agent review WANDR, and Zapier's automation review AutomationBench, costing less than half the cost per task of the second-place external model. Google didn't name which model came in second place.

Google has also cooperated with the evaluation agency Artificial Analysis to launch a new multi-modal knowledge agent evaluation MKA, which tests the model's ability to read dense documents such as diagrams, blueprints, tables, etc. Carbon ranked second in this review, with an average cost per task that was less than one-sixth of the cost of the first place.

Antigravity simultaneously previews the /swarm command, which can be split into a set of agents running simultaneously with a single prompt. Developers can hand over the most reasoning-intensive roles such as architecture design, coordination, and auditing to Carbon, and use faster and lighter models for other tasks.

Wiz used Carbon to find zero-day vulnerabilities in two large SaaS companies in two hours

Information security is Carbon’s main capability, and the enhanced version of information security is still open for application only through Google’s Fairwind program. Google cited several early test cases: Wiz, a security company, used Carbon to check two large SaaS companies in a public interest scanning program. It found zero-day vulnerabilities in less than two hours and has completed patching with the affected companies.

The Google Cloud Vulnerability Research Team also used Carbon to write a proof of concept, proving that two bugs in a popular open source software could indeed be exploited. The severity was comparable to the recent PHP2Shell vulnerability, so the maintainer classified it as a major vulnerability and released a patch. Google said details will be released after the responsible disclosure process is completed.

Honesty test passed 92%, Argon 70%

In terms of security, Google said that Carbon passed 92% of the 375 agent honesty tests, and Argon passed 70%; when encountering experimental data gaps, it refused to make up for it on its own in 98% of the cases.

In tests simulating a network attack, red team agents' breach rate per attempt dropped from Argon's 3.1% to 0.8%. The success rate of automated attack queries measured by the external agency Gray Swan was 2.17%, and that of Argon was 9.86%; Trajectory Labs spent 132 hours and sent more than 13,000 requests, but did not find a new universal jailbreak method.

Google also admitted that there is still room for improvement in prompt injection protection, and its own automated red team testing showed that the model still has room to be strengthened in the face of some more sophisticated attacks. For complete data, please see Google’s announcement Carbon Security Report。

this article Gemini 4 Carbon is open to paying developers, and the API price will double starting in December first appeared in Chain News ABMedia。