geminiai-modelsmodel-selectionai-costsapi-pricing

Gemini 4 Carbon for business: on sale, pricier from 1 Dec

November 12, 20267 min readPixel Management

This article is also available in Dutch

On 10 November 2026 Google opened Gemini 4 Carbon to every paying business. Two weeks earlier it was reserved for a hand-picked group of cyber defenders.

Gemini 4 Carbon for business has been available since 10 November 2026 through the Gemini API and Gemini for business, at an introductory price of $2 per million input tokens and $10 per million output tokens until 30 November. From 1 December it costs $4 / $20, exactly what Gemini 4 Argon and Claude Opus 5.5 will cost on that date too. For a Dutch company that means two things: you have eighteen days to test cheaply, and once they're over, price stops being a reason to pick Carbon or skip it.

What changed on 10 November?

From 28 October Gemini 4 Carbon was only available to "a subset of cyber defenders" through Google's Fairwind Program, as Google wrote when it introduced the model. General availability followed on 10 November: Carbon is now in Google AI Studio and the paid Gemini API for "all paying developers and enterprises", in Gemini for business, and in the Gemini app for Google AI Ultra subscribers.

There's no free tier. The Gemini API pricing page lists Carbon as "Not available" in the free column. On the paid tier Google doesn't use your inputs and outputs to improve its products, which the same page states per model. Google Workspace isn't mentioned in the announcement.

Google describes Carbon's purpose as "sustained agentic workflows": long, multi-step tasks involving code, search and tools. At launch Google claimed first place on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at "more than 2X lower cost per task compared to the next best external model". Those are the vendor's numbers, on tasks the vendor picked. How to weigh a claim like that is in evaluating AI benchmarks.

What does Gemini 4 Carbon cost before and after 30 November?

Until 30 November 2026 Gemini 4 Carbon costs half of what it costs afterwards. The doubling applies to every rate on the pricing page, caching and batch included.

Rate per million tokensUntil 30 Nov 2026From 1 Dec 2026
Input$2.00$4.00
Output, including thinking tokens$10.00$20.00
Cached input$0.10$0.20
Batch, input / output$1.00 / $5.00$2.00 / $10.00

Two details matter when you read that table. Thinking tokens are billed as output, the expensive rate, and a reasoning model like Carbon produces a lot of them. And Gemini 4 Argon doubles on the same day, from $2 / $10 to $4 / $20. If you run Argon today, you get the price rise either way. Everything else that changes on 1 December is in our AI model retirement calendar.

Run it on a real volume. A process that uses 500 million input tokens and 100 million output tokens a month costs 500 × $2 + 100 × $10 = $2,000 on Carbon in November. In December it's 500 × $4 + 100 × $20 = $4,000. Those figures are indicative, with no caching or batch discount, and they show why a business case built on the introductory rate is off by a factor of two.

Is the introductory price a reason to switch now?

The introductory price is a reason to test now, not to switch now. A proper test is cheap at either rate, and a production decision you make because November is cheap gets billed at double from December.

Here's what a test costs. Take a test set of 200 real cases from your own process, each run three times because a reasoning model doesn't answer the same way every time. Assume 20,000 input tokens and 4,000 output tokens per case, thinking included. That's 600 runs: 12 million input tokens and 2.4 million output tokens. In November it costs 12 × $2 + 2.4 × $10 = $48. In December, $96. A $48 difference doesn't justify rushing. What does count: in November you can run Carbon next to Argon and measure the gap while both are still cheap, before you lock in a 2027 budget in December.

Time saved

Save 6 hours per week on manually comparing AI models at every new release

From 1 December Carbon, Argon and Claude Opus 5.5 cost exactly the same per token: $4 / $20. Token price then decides nothing. What's left is how many tokens each model needs for your task, and how often it gets the task right. That's cost per task, and no pricing page shows it. You measure it yourself, or have it measured.

Your situationWhat to do before 30 November
You run Gemini 4 Argon in productionRun your test set on Carbon and Argon, compare cost per correctly finished task
You run Claude or GPT and are considering GoogleTest Carbon now, decide after 1 December at the standard rate
You process personal dataTest through Google Cloud in the EU multi-region, not through AI Studio
You don't have a test setBuild one first; without it you're measuring the wrong things

Does Gemini 4 Carbon run in the EU?

Gemini 4 Carbon runs in the EU if you call it through Google Cloud in the EU multi-region. According to Google Cloud's data residency page, Google guarantees machine-learning processing inside the EU for Carbon only in that multi-region. No single European region, such as europe-west4 in Eemshaven, is ticked for Carbon.

The global endpoint gives no such guarantee. Google writes that requests there "may be processed in any Google Cloud location around the world, and therefore don't provide any data residency guarantees". For a test with made-up examples that doesn't matter. For a test with real customer files it does: the endpoint choice is then a GDPR decision, and it should be settled before the first request. What EU data residency with a US provider does and doesn't solve is covered in EU data residency with US AI providers.

Note that the prices in the table above are Gemini API prices. Google Cloud bills its Agent Platform (formerly Vertex AI) separately, so check the rate for your chosen region before you set a budget.

When is Gemini 4 Carbon the wrong choice?

Gemini 4 Carbon is the wrong choice for simple, high-volume work. A model built for long agent tasks also thinks hard about an invoice line, and you pay for that thinking as output. A small model classifies, summarises or pulls fields out of a document for a fraction of the price.

Agents that read web pages, emails or documents on their own need extra care. Google reports a 0 percent attack success rate on the Gray Swan indirect prompt injection benchmark, and in the same announcement adds that "prompt injection robustness remains an area for continued improvement". Take that literally. An agent that sends email or makes payments on your behalf is secured in the architecture, not by the model choice; how is in securing AI agents against prompt injection.

For a decision wider than one model, the buying side is in our guide to hiring AI consulting. Anthropic released its own top model a day earlier, on different terms; we cover it in Claude Fable 5.2 for business. If you want Carbon tested on your own process before 30 November, with a test set you keep for every release after it, we do that as part of AI consulting.

Service

Learn more about AI consulting?

View service →

Curious how much time you could save?

Request a free efficiency audit. We'll analyze your processes and show you where the gains are, no strings attached.