Google Gemini 4 Carbon Is Now Open to Paid Developers — Price Doubles on 1 December
Google has opened Gemini 4 Carbon to paid developers, and the price on offer today will not last long. From 10 November 2026, paying developers can use the model through the Gemini API in Google AI Studio and Android Studio, at $2 per million input tokens and $10 per million output tokens. That introductory price runs through 30 November 2026, and from 1 December 2026 the price doubles to $4 input and $20 output per million tokens. Gemini 4 Carbon is Google's heavy agentic model, built for long, multi-step work that cheaper models handle less reliably, and until now it was available only to a narrow group of cyber defenders.
What is Gemini 4 Carbon
Gemini 4 Carbon is the top tier of Google's Gemini 4 family, aimed at agentic work. An agentic model is one that plans a task, takes a series of steps toward it, uses tools along the way, and checks its own progress, rather than simply answering a single question in one pass. Google positions Carbon above Gemini 4 Argon, which remains the recommended model for routine work. Carbon is the model for long, complex, or high-stakes tasks, such as a multi-hour coding job or an automation workflow that must keep going without constant human steering. Readers who want the full picture can read our flagship AI model comparison, and our first coverage of the model is in our report on Gemini 4 Carbon.
The model first appeared on 28 October 2026, but only for a limited audience: Google made it available then to cyber defenders through its Fairwind programme, a restricted early-access route for security teams. That limited start is now over. The 10 November 2026 announcement on the Google blog, titled "Gemini 4 Carbon is now available", extends access to developers, enterprises, scientists, and consumers, with the exact route depending on who you are and what you pay for.
Who can use it, and where
For developers, access runs through the Gemini API paid tier in Google AI Studio and Android Studio, and Google says it is open to all paying developers there. There is no free API tier for Carbon, so developers on the free tier will need to upgrade to a paid plan before they can call the model.
Enterprises get access through Gemini for Business. Consumers on the top subscription tier are also included, though on a rollout rather than an immediate switch-on: Google says Carbon is rolling out to all Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search, so some Ultra subscribers will see it before others, and Google has not published a date by which every Ultra subscriber will have it.
Gemini 4 Carbon pricing: what it costs now, and from 1 December
The pricing is the part of this announcement that most directly affects developers, because the introductory rate is time-limited. Through 30 November 2026, Carbon costs $2 per million input tokens and $10 per million output tokens. From 1 December 2026, those rates double to $4 per million input tokens and $20 per million output tokens. Cached input costs $0.10 per million tokens through 30 November and $0.20 per million tokens from 1 December. Thinking tokens, the tokens the model uses for its internal reasoning before it produces an answer, are billed as output tokens, which means long reasoning chains are charged at the higher output rate.
| Price item | Through 30 Nov 2026 | From 1 Dec 2026 |
|---|---|---|
| Input, per 1M tokens | $2 | $4 |
| Output, per 1M tokens | $10 | $20 |
| Cached input, per 1M tokens | $0.10 | $0.20 |
| Thinking tokens | Billed as output | Billed as output |
The doubling matters for budgeting: a developer who builds a workflow around Carbon in November at the introductory rate will see the same usage cost twice as much in December. Because output tokens cost five times as much as input tokens, and thinking tokens are billed as output, tasks that make the model reason at length will be the expensive ones. Caching helps where the same context is reused, since cached input is far cheaper than fresh input at either price point. November is the month to measure real usage before the December price takes effect.
 |
| A software developer working at a desk with code on multiple monitors in a dark modern office at night |
Gemini 4 Carbon benchmarks, as reported by Google
Google says Carbon ranks number one on three benchmarks: CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench. All of these results are Google-reported, and there has been no independent check of them. One caution is important here. VentureBeat found no Carbon row on the public CursorBench and WANDR tables, so this article gives no scores for those two benchmarks and describes the number-one claims as Google's own statements. Readers should treat them as vendor-reported results until an independent table or a public leaderboard confirms them.
The one benchmark where Google has published full numbers is Zapier AutomationBench. On AutomationBench, Carbon scored 53.9 percent on High effort tasks, against 51.3 percent for Argon, 44.7 percent for Claude Sonnet 5.5, and 41.4 percent for GPT-6 Astra. The test covered 657 tasks on version 1.0.6 of the benchmark. The gap between Carbon and its closest rival, its own sibling Argon, is 2.6 percentage points, while the gap to Claude Sonnet 5.5 is 9.2 points and to GPT-6 Astra is 12.5 points. For readers who want the rival's numbers in full, our report on GPT-6 Astra benchmark scores covers that model in detail.
| Model | AutomationBench, High effort |
|---|---|
| Gemini 4 Carbon | 53.9% |
| Gemini 4 Argon | 51.3% |
| Claude Sonnet 5.5 | 44.7% |
| GPT-6 Astra | 41.4% |
 |
| Close-up of hands typing code on a laptop keyboard with a glowing screen in a dim room |
Gemini 4 Carbon vs Argon: which one to use
The choice inside the Gemini 4 family is now a clear one, and Google has made its own recommendation. Argon remains the recommended model for routine work: the everyday coding, writing, and question-answering tasks where a fast, cheaper model is enough. Carbon is the heavy agentic tier, for the tasks where Argon would struggle or need more hand-holding. The AutomationBench numbers give that recommendation a measurable shape: on High effort automation tasks, Carbon beats Argon by 2.6 percentage points, a modest gap, but on the hardest tasks in the test, where a few points can mean the difference between a workflow that finishes on its own and one that stalls and needs a human to rescue it.
Price is the other half of the decision. Carbon at $2 input and $10 output through November, and $4 and $20 from December, is priced as a premium model. The sensible approach is to use Argon by default and reach for Carbon where the task is long, multi-step, or high-stakes, and to measure the difference on real work rather than assuming the premium model is worth it everywhere. The November window is useful for exactly that test: run the same workflow on both models at the introductory Carbon price, compare the success rate and the cost, and decide before the December doubling whether Carbon earns its place in production.
 |
| A modern data centre corridor with rows of server racks and blue LED light |
Key Takeaways
- Gemini 4 Carbon is now open to all paying developers through the Gemini API in Google AI Studio and Android Studio, with no free API tier.
- It costs $2 input and $10 output per 1M tokens through 30 November 2026, doubling to $4 and $20 from 1 December 2026; cached input is $0.10, then $0.20, and thinking tokens are billed as output.
- Google reports Carbon first on CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench, with 53.9 percent on AutomationBench High effort, ahead of Argon at 51.3 percent, Claude Sonnet 5.5 at 44.7 percent and GPT-6 Astra at 41.4 percent.
- VentureBeat found no Carbon row on the public CursorBench and WANDR tables, so those two number-one claims are Google's word only, with no scores given here.
- Argon stays the recommended model for routine work; Carbon is the heavy agentic tier, and enterprises get it through Gemini for Business while Ultra subscribers receive it on a rollout.
Conclusion
Gemini 4 Carbon is now a real option for paid developers rather than a restricted preview, and the terms are clear: $2 and $10 per million tokens until the end of November, then double from 1 December, with no free tier. Google's benchmark claims put it ahead of Argon, Claude Sonnet 5.5, and GPT-6 Astra on AutomationBench, but the missing public rows for two of the three claimed number-one finishes mean developers should test the model on their own work before trusting the headline. Use November to measure it, compare it with Argon, and decide before the price rise whether the heavy agentic tier earns its cost.
Comments
Post a Comment