Gemini 4 Carbon: release date, price, benchmarks and how it differs from Argon

Google announced Gemini 4 Carbon on 28 October 2026 and calls it its strongest model, but at first only a subset of cyber defenders can use it. Developers, businesses and consumers get it later, on a date Google has not given, at Argon's launch price of $2 per million input tokens and $10 per million output tokens.

So for now Carbon is a model you plan for, not one you call. Gemini 4 Argon, generally available since 13 October, is still the Gemini 4 model in the API. Below: who gets Carbon first, what Google says it does better than Argon, and what is still unknown.

Gemini 4 Carbon release date

There is no release date for developers yet. Google's rollout runs in 3 stages, and only the first has started:

1. The week of 28 October: a subset of cyber defenders.
2. Then, gradually: more trusted partners, through Google DeepMind's Fairwind Program.
3. Later: developers, businesses and consumers, "as soon as possible". Google says it is still gathering feedback and working on guardrails first.

Google also says it takes part in the US government's voluntary testing of models before release. On 28 October Carbon was not in the Gemini API release notes, and Google had published no model id, so no code can name it yet.

Fairwind is not open to everyone. Its page puts governments and national cyber authorities, critical infrastructure operators and core technology platforms first, and vets each applicant's record of ethical operations. If your team defends systems in one of those groups, the interest form on that page is the way in. On 28 October the page still described its access as Gemini 4 Argon with cyber defence capabilities.

Gemini 4 Carbon and Argon compared

Carbon builds on Argon, which launched 2 weeks earlier. It shares Argon's safety stack (training, real-time misuse defences, resistance to prompt injection and monitoring of the model's chain of thought) and its launch price. Where Google compares them directly:

Every Carbon figure here is Google's own. Nobody outside Google has published a test of Carbon that the reader can check, and the internal tests are not public.

The prompt injection row matters most if you build agents. An attack success rate of 0.7% sounds small, but an agent that reads 1,000 web pages or emails a day gets 1,000 chances to be steered. Google's 0.0% is on one benchmark's set of attacks; treat it as a better lock, not a reason to give an agent more access. The prompt injection page covers the controls that stay necessary whatever the model.

Gemini 4 Carbon benchmarks

Google's announcement leads with cyber defence, the job Carbon goes to first. It reports these scores, and says the outside benchmarks were run without guardrails, the guardrails it is still working on before a wider release:

Google also reports how Carbon behaves on long jobs. Its teams have seen it run for days on multi-step engineering work, such as finding the cause of subtle bugs across large codebases and patching them. On Terminal-Bench 4.0 it cheated successfully 0.0% of the time and tried to in 0.4% of tasks. In Google's live coding work, high-severity misalignment alerts fell by almost 80%, overreach by 26% and under-service by 29%.

A ranking from Google's own launch post is a starting point, not a verdict. When Carbon reaches the API, the scores that matter are on your own prompts: the guide to evaluating an AI agent shows how to run that comparison, and the eval sample size calculator tells you whether a difference is real.

What Google has not said yet

The announcement leaves out what a developer needs before writing code:

• A model id. Neither the announcement nor the Gemini API release notes give one.
• Context window and output limit. Argon has a 1,048,576-token window shared by input, thinking and output. Google has not said whether Carbon matches it.
• When the launch price ends. Argon's offer ends on 30 November 2026. Carbon's has no date.
• Where it runs. The post names no product: not Google AI Studio, not Google Cloud, not the Gemini app.

What to build on until Carbon arrives

Build on Argon. Carbon builds on it and launches at the same price, and Argon is the Gemini 4 model you can test your prompts, tools and settings on today.

2 things to plan for now:

• The price step. Argon's rates double on 1 December 2026, to the same $4 and $20 Carbon will reach after its own offer. Price your workload at the later rate; the flagship comparison works through what that does to a month of jobs against Claude Opus 5.5 and GPT-6 Astra.
• The switch itself. When the id appears, changing one model string is the easy part. Keep the model name in configuration, not code, and run your own test set on both before you move traffic. If Carbon turns out to be more than most requests need, model routing shows how to send only the hard ones to the stronger model.

We add Carbon's price, window and model page to /prices once Google publishes them, and the change goes on the What changed timeline.