Published: 11 November 2026
Google opened Gemini 4 Carbon, its strongest model, to all paying Gemini API developers and businesses on 10 November, two weeks after it went only to cyber defenders. Until 30 November it costs the same as Gemini 4 Argon: $2 per million input tokens (small chunks of text) and $10 per million output tokens. From 1 December that becomes $4 and $20, and Google itself says Argon is still the better choice for most work.
Developers get Carbon in AI Studio, consumers only with AI Ultra
Developers reach Carbon through the paid tier of the Gemini API in Google AI Studio and Android Studio, under the model name gemini-4-carbon. Companies get it through Gemini for business, via the API or in the model menu, and in Google’s coding tools Antigravity and CodeMender.
In the Gemini app and AI Mode in Search, Google is rolling Carbon out to Google AI Ultra subscribers. The announcement does not mention the cheaper AI Pro plan.
Two smaller groups get special access. Scientists can apply for six months of free use of Argon and Carbon in the highest usage tier of Antigravity. Security teams that want a version with fewer safeguards still go through Google’s Fairwind Program, where Carbon first appeared in late October.
Carbon bills the same per token as Argon, but it uses more tokens per job
Google writes that Carbon spends more tokens than Argon “to explore thoroughly, reason deeply, and run extended verification loops”. The same price per token therefore does not mean the same bill. The Decoder made this point on launch day: the total cost of a task on Carbon is likely higher than on Argon.
Google is also clear about the split between the two. Argon remains “the ideal choice for the majority of software engineering and knowledge work”, while Carbon is built for tasks that “test the limits of a model’s intelligence”. In practice that means long agent runs (AI that works through many steps on its own), security reviews of a codebase, and analysis where checking the answer matters more than speed.
One result stands out for knowledge work. On a test of 375 tasks that Google built itself, Carbon gave an honest answer in 92% of cases, against 70% for Argon, and declined to make up missing data 98% of the time. No outside party has checked that test yet.
Fable 5.2 scores higher, but Carbon costs less per task
The figures below show Carbon giving up a few points of capability against Claude Fable 5.2 in return for a much lower bill.
| What is measured | Gemini 4 Carbon | Claude Fable 5.2 |
|---|---|---|
| Input price per million tokens | $2 until 30 November, then $4 | $8 |
| Output price per million tokens | $10 until 30 November, then $20 | $40 |
| Artificial Analysis Intelligence Index | 60 (high effort) | 62 (high), 65 (max) |
| Cost per task on that index | $7.76 (high) | $9.08 (high), $26.54 (max) |
| Zapier AutomationBench | 53.88% (high) | 51.45% (max) |
Prices come from Google and Anthropic. The index and cost figures come from Artificial Analysis and the AutomationBench scores from Zapier, both as reported by VentureBeat on 10 November. Each model ran at its own effort setting, so read the rows as a guide rather than a strict ranking.
On the same index, OpenAI’s GPT-6 Astra is cheaper still at $6.24 per task, with a score of 57. Google’s post claims first place on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at less than half the cost per task of the next best outside model. It gives no scores. Zapier’s numbers back the AutomationBench claim. VentureBeat found no Carbon row in the public CursorBench table, and Anthropic’s WANDR table leaves Carbon blank.
The Netherlands is on Google’s API list, but Google names no data region for Carbon
Google’s announcement does not mention Europe. Its list of countries where the Gemini API and AI Studio work does include the Netherlands and every other EU country, so Dutch developers on the paid tier can call Carbon today. Google has not published a separate country list for Carbon.
Google also does not say where Carbon processes data for Gemini for business customers. If your contracts require processing inside the EU, ask your Google account team before you move production work to Carbon.
Three weeks at Argon’s price are enough to find where Carbon pays off
For most organisations DataNorth works with, Carbon is not a model to move everything to. Everyday drafting, summarising and coding run well on Argon or smaller models, and Google says the same. Carbon earns its place in a few spots: agents that run many steps without a person checking each one, security review of your own code, and analysis where one invented number does real damage.
Use the low price as a test window. Take ten to twenty real tasks, run each on Argon and on Carbon, and log the success rate and the tokens used per task. Then price the results at $4 and $20, not at today’s rate. If Carbon only wins on a handful of tasks, send those to Carbon and keep the rest on Argon. Teams on Gemini for business should also agree who may pick Carbon from the model menu, before colleagues use it for routine work.
For more information, visit the official announcement of Gemini 4 Carbon on the Google blog.
