Gemini 4 Carbon is now public at $2 per million tokens. What it changes for your business

On 10 November 2026, Google made Gemini 4 Carbon generally available. Google's most capable model is now in the paid Gemini API, in Gemini for business, and in Google AI Ultra subscriptions. The launch price is $2 per million input tokens and $10 per million output tokens, through 30 November. After that it rises to $4 and $20. For comparison, Claude Fable 5.2, released a day earlier, costs $8 and $40. The question for a business is not which model wins on paper, but which model finishes your jobs at a price that works.

What exactly launched

Carbon was announced on 28 October and first offered only to selected cybersecurity defenders through Google DeepMind's Fairwind program, for work such as finding software vulnerabilities. 10 November is the wider launch of the general-use model.

Google has not published a country-by-country rollout schedule, and availability in Vertex AI was also left unspecified in the launch materials. If your firm is in Germany or Canada, check availability in your region before making plans, instead of assuming the launch price applies everywhere.

The prices, worked out properly

The introductory price of $2 and $10 per million tokens runs through 30 November 2026. After that it doubles to $4 and $20. Even after doubling, it is half the price of Claude Fable 5.2 at $8 and $40.

One worked example: one million input tokens and 200,000 output tokens cost $4 with Carbon during the introductory period, and $8 after 30 November. The same job costs $16 with Fable 5.2. Google had also announced a 95% discount on cached input, which makes jobs with repeated content cheaper still.

Price is only half the bill, though. What a model costs per completed job depends on how often it fails a job and has to repeat it. A cheaper model that fails nearly half its automated tasks still bleeds budget on retries.

The benchmarks are split

The rankings split by workload, and that is exactly what you should use.

Carbon scores 60 on the Artificial Analysis Intelligence Index at high reasoning. Claude Fable 5.2 leads there at 65 with maximum effort, and 62 at high effort. Claude Opus 5.5 sits at 59, GPT-6 Astra at 57. Carbon debuted in first place on this index at its October announcement. Fable has since taken the lead.

Carbon leads in two practical tests. In Zapier's AutomationBench, Carbon scores 53.88%. In Terminal-Bench 4.0, Carbon scores 65.2%, ahead of Fable at 61%. Artificial Analysis puts Carbon's cost at $7.76 per task, against $26.54 per task for Fable at maximum effort. So Carbon is cheaper in practice for automation and coding work, while Fable leads the general intelligence ranking.

How to decide what to test

Do not choose by the index. Choose by your three most important jobs.

Step one: write down three jobs your team does with AI every week. A long research task, a coding job, and an automation between your business apps are a good start.

Step two: run the same job once on Carbon and once on the model you use today. Count the cost per job, not per token, and count how often you had to fix the answer.

Step three: decide by completed jobs. If Carbon finishes your automations cleanly, it is the cheaper choice for that work. If Fable solves the harder research better, keep it for that one job and do not switch your whole stack.

Step four: mind 30 November. The introductory price doubles then. Test before that date, so you know whether $4 and $20 still works for you.

What this means for German and Canadian businesses

The price war is good for small businesses. Two frontier models competing for the same customer means lower prices and faster releases. But it also makes choosing harder. A model that leads today can be second again next week. Do not build your workflow on one model. Build it on the job: what the AI has to do, what a finished job costs, and who checks the result.

For teams that do a lot of automation work, Carbon is worth the test. For teams with hard research or large code refactors, Fable 5.2 stays the stronger candidate. Read our blog on Claude Fable 5.2 for the other side of the comparison.

The takeaway

Gemini 4 Carbon has been public since 10 November 2026, at $2 per million input tokens and $10 per million output tokens through 30 November, then double. It costs a quarter of Claude Fable 5.2, leads in automation and coding tests, and trails in the general ranking. Test it on your three most important jobs, count the cost per completed job, and decide before 30 November, when the price doubles.

Sources: VentureBeat, 10 November 2026; Zubiqo, based on VentureBeat, 10 November 2026.