Google opens Gemini 4 Carbon to all paid developers, and the introductory price doubles after 30 November
Listen to this article
Read by Anchor
About two weeks ago Gemini 4 Carbon was limited to a small group of cyber-defense researchers. On 10 November 2026 that changed. Koray Kavukcuoglu, vice president of Google DeepMind and senior AI engineer at Google, wrote: “Today we are making Gemini 4 Carbon available to developers, enterprises, researchers and consumers.” However the release came with a pricing tier that actually started to circulate, and that is where the story begins.
Who gets it today and who waits
The announcement page says the model is now available through the paid tier of the Gemini API within Google AI Studio and Android Studio, and via Gemini for Business for enterprises. Google AI Ultra subscribers receive the model gradually in the Anti-Gravity environment, the Gemini app and the AI-powered Google Search. The company also opened applications for researchers to obtain six months of free access to Argon and Carbon through the highest usage tier in Anti-Gravity, subject to eligibility, capacity and selection. Specialized cyber access remains through the Fairwind Program in successive batches.
The price that changes after 30 November
The introductory price is $2 per million input tokens and $10 per million output tokens, the same as Gemini 4 Argon. A token is the smallest unit of text that developers are billed for. After 30 November 2026 the price becomes $4 for input and $20 for output, meaning the bill doubles twenty days after the announcement. Anyone who built their calculations on the introductory price must recalculate now, not in December.
Numbers Google tells about itself
Google notes that the model first appeared in three programming-automation agent tests, CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench, at a per-task cost more than twice lower than the best competing external model. It placed second in the multimodal knowledge test MKA at a median cost more than six times lower than the first model. On the security side, Google says the breach rate per attempt fell to 0.8% from 3.1% in Argon, and the model passed 92% of the agent truthfulness benchmark’s 375 tasks versus 70% for its predecessor. All these figures come from Google itself, and we have not found independent verification. The page acknowledges that the model’s robustness to prompt injection remains an area for improvement.
What the page says and what it omits
The lens here is sovereign. In the 28 October announcement Google said Carbon was part of the voluntary pre-launch model testing for the US government, and access was rolled out in phases: a batch of cyber-defense researchers first, then trusted partners via Fairwind, then developers, companies and consumers. Google set the phase order, and the page does not mention our region’s location. The 10 November page does not state usage limits or availability by country. Therefore we cannot say that a developer in Riyadh, Abu Dhabi or Cairo runs the model today under the same conditions as a developer in San Francisco.
Practical takeaway
Three points merit attention. First, pricing will double at the end of the month, so the cheap-trial window is short. Second, Google’s tests measure what Google chose to measure, and the prudent judgment is to run the model on your own tasks before moving any product to it. Third, anyone building on a model whose rollout phases are set outside the region needs a ready alternative, because access dates may shift by a decision beyond their control. On the positive side, the announcement does have a clear benefit: researchers who meet the criteria can now apply for six months of free access, and applications are open according to the page.