Google opens Gemini 4 Carbon to paying developers, enterprises and AI Ultra subscribers

Google made its Gemini 4 Carbon model available on 10 November 2026 to developers and enterprises on paid plans, and began rolling it out to subscribers of its AI Ultra plan. Koray Kavukcuoglu, Google DeepMind's SVP and Google's Chief AI Architect, announced the release in a blog post.

Paying developers reach Carbon by API through Google AI Studio, and enterprises through Gemini for business. AI Ultra subscribers are getting it across Google Antigravity, Search's AI Mode and the Gemini app. Google's pricing page says the model is not offered on the free tier, and that every token it processes, thinking tokens included, is billed.

Price and access

The standard rate is $2 for input and $10 for output, per million tokens, through 30 November 2026. From 1 December 2026 the rates become $4 and $20. Google's post describes the lower figures as an introductory price, the same one it set for Gemini 4 Argon. The pricing page also lists cheaper batch and flex tiers.

Carbon was first announced on 28 October 2026. Google said then that it would give the model to a limited group of cyber defenders and widen access gradually to trusted partners through its Fairwind Program, while it gathered feedback and refined its guardrails. It also said it was participating in the US government's voluntary process for testing models before release. Defenders can still apply to Fairwind.

What Google says it can do

Google calls Carbon its strongest model. It says Carbon is meant for work that pushes a model's limits, such as complex coding and debugging over many steps, and exploratory scientific research, and that it spends more tokens than Argon so it can explore, reason and check its own work. Google says Argon remains the better choice for most knowledge work and software engineering.

The company says Carbon ranks first on Zapier's AutomationBench, Perplexity's WANDR and CursorBench 4.0, at less than half the cost per task of the next best model from outside Google. On a new Multimodal Knowledge Agent benchmark, launched with Artificial Analysis, Google says Carbon placed second at under a sixth of the leading model's average cost per task. In the October announcement it reported a score of 59 on version 5.0 of the Artificial Analysis Intelligence Index, and 88.2 percent on SEC-bench Pro, 54.8 percent on ExploitGym and 71.0 percent on CWE-bench for finding, exploiting and patching software flaws. All are Google's own figures.

Google is also previewing a /swarm feature in Antigravity, which turns one prompt into a coordinated group of agents. On safety, it says Carbon blocked 99 percent of attack paths from its adaptive red-teaming agent in simulated cyberattack tests, and that its breach rate per attempt fell to 0.8 percent from 3.1 percent on Argon.

Google is also extending Carbon to scientists in a programme, opened in October, that offers 10,000 of them six months of free access to Antigravity. Applications remain open, with weekly onboarding that depends on eligibility, capacity and selection.