Alphabet’s Google on Tuesday widened access to Gemini 4 Carbon, the most capable model it has built, moving the system out of a two-week restricted trial with a handful of cybersecurity firms and into the hands of paying developers worldwide, according to the company.
The model is now callable through the API in Google AI Studio and Gemini for business. Subscribers to Google AI Ultra, the $100-a-month tier, can also reach it inside the Gemini app, AI Mode in Search and Antigravity, the company’s agentic software development platform. Pro-tier and free users are excluded for now.
Google is holding introductory API pricing at the same level as the prior Gemini 4 Argon model: $2 per million input tokens and $10 per million output tokens. That rate expires Nov. 30, after which the company said the price doubles.
| Gemini 4 Carbon at launch | Detail |
|---|---|
| API input price (introductory) | $2 per million tokens |
| API output price (introductory) | $10 per million tokens |
| Price change | Doubles after Nov. 30 |
| Consumer access | Google AI Ultra subscribers only ($100/month) |
| Developer access | API via Google AI Studio, Gemini for business |
| Restricted variant | Carbon for cyber, via Fairwind program application |
Google positioned Carbon as a tool for hard problems rather than everyday chores. The company said the model suits complex programming, deep debugging and scientific research, and that it consumes more tokens because it reasons and verifies its output repeatedly. For routine work, Google continues to recommend Argon, the first Gemini 4 model and a less advanced system.
The launch came with a security demonstration. Wiz, the cybersecurity company Google owns, ran Carbon against two large software vendors and surfaced previously unknown zero-day vulnerabilities in under two hours, then helped the affected companies patch them. Google said work of that kind has typically taken weeks. The caveat: Wiz was using Carbon for cyber, a specialized edition that still requires an application through the Fairwind program. The version released this week is the general-purpose model.
Benchmark results released alongside the launch put Carbon at the top of several third-party leaderboards, including CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, while Google claims cost per task more than 2X lower than rival external models. On security-specific tests such as SEC-bench Pro, ExploitGym and CWE-bench v1, the company said the model can identify, exploit and remediate vulnerabilities.
Google also said Carbon has shown gains in protein structure prediction, a step in drug development, and that it is granting 10,000 scientists six months of free access to support their research. Thousands have already been onboarded, the company said.
Multimodal work is part of the pitch as well. Google and Artificial Analysis built the Multimodal Knowledge Agent benchmark, on which Carbon placed second while keeping average cost per task roughly 6X below the top model. Using Antigravity’s /swarm feature, the company demonstrated the model coordinating multiple agents to rebuild a 3D model of the Eiffel Tower’s construction from historical photographs and engineering plates, and to assemble an interactive textile visualizer from academic research.
For Google, the release sharpens a two-tier product line: Argon for volume work, Carbon for high-precision tasks in legal and financial analysis, full-stack development and 3D design. The Nov. 30 price step gives developers a narrow window to lock in work at the lower rate, and the gated cyber edition keeps the most sensitive capability behind an application process even as the general model goes wide.