Gemini 4 Carbon is out at $2 in and $10 out per million tokens until November 30, and it is already in n8n. Price it per task, not per token
A new frontier model at a low per token price is good news, as long as you budget for the price it will have in three weeks.
What happened
Google made Gemini 4 Carbon generally available through the Gemini API in Google AI Studio and Gemini for business, for paying developers and enterprises, after two weeks with a select group of security defenders. Google says Argon is still the right choice for most software engineering and knowledge work, while Carbon is for harder multi-step coding, debugging and research.
The introductory price is $2 per million input tokens and $10 per million output tokens. n8n says it rises to $4 and $20 after November 30, 2026. The Verge notes that cheaper tokens don't automatically mean cheaper work, and The Decoder reports Carbon uses more tokens than Argon. In Artificial Analysis testing at the standard price, Carbon cost $7.76 per solved task.
n8n added Carbon the same day. It cites Google's results: 53.9% on Zapier's AutomationBench and a 0.0% success rate for attacks on Gray Swan's prompt injection benchmark, down from 0.7% for Argon. n8n still says to keep security protocols in place. Carbon is available in the AI Agent node through the Google Gemini Chat Model, and on n8n Cloud it runs on Gateway credits. Google also previewed a swarm feature in Antigravity that turns one prompt into a team of agents.
My take
Here is how I'd test it in a client workflow:
- Price the task, not the token. Run 20 real executions on Carbon and on your current model, and compare total cost per finished job. A model that thinks longer can cost more even at a lower rate.
- Budget at $4 and $20. Anything you ship this month will still be running in December.
- Use it where reasoning pays. n8n's own advice fits: Carbon on the hard agent steps, Argon or a lighter model for classifying and formatting.
- Keep your injection guards. Inbound emails and webhooks are still untrusted. A great benchmark score is not a reason to give an agent write access without checks.
Swap the model behind one agent step, measure it, and only then roll it out.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
