Gemini 4 Carbon: Google's strongest model, but who can use it, and what will it cost?

On 28 October 2026 Google announced Gemini 4 Carbon. Google calls it "our strongest model", made for computer security and for long jobs with many steps. It comes about two weeks after Gemini 4 Argon opened to paying developers. But most people cannot use it yet. Google says it will first give Carbon to a small group of cyber defenders, people who protect computer networks. It gives no date for anyone else.

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google's newest top model. The announcement was written by Koray Kavukcuoglu, a senior vice president at Google DeepMind. He says Carbon builds on Argon. It is meant for "sustained agentic workflows": long jobs where the AI works through many steps by itself.

Google gives one example. Its own teams saw Carbon "run for days" on hard engineering jobs. One example is finding the cause of small bugs in large programs, and fixing them. We could not check that ourselves, so we report it as Google's claim.

A token is a small piece of text. AI companies count what you send and what you get back in tokens. That is how they set the price.

Who can use it, and when?

Google says access starts with selected cyber defenders. It writes that it will make Carbon available "to a subset of cyber defenders" starting this week. After that it will "gradually phase access to additional trusted partners" through its Fairwind Program. We could not check who already has access.

Developers, businesses and normal users come after that. Google says it will bring Carbon to them "as soon as possible". That sentence has no date and no region. It does not name the Gemini app or the API either.

Google calls this "a deliberate approach". It also takes part in the U.S. government's voluntary testing of new models before they come out. Argon took the same road: Google announced it on 30 September and opened it to paying developers on 13 October.

What will it cost?

Google gives the price in its post. Carbon starts at the same price as Argon: $2 per million tokens in and $10 per million tokens out. Cached input is text saved so the model can use it again. Google says those input tokens cost 95% less. A small note under the post adds the later price: $4 in and $20 out.

Google does not say when the starting price ends. For Argon, Google's price list says the starting price runs until 30 November 2026.

Prices are per million tokens, from each company's own page on 29 October 2026. But the price per token is only half the story. A model that writes more costs more per job. In our test, Argon wrote few tokens, so a coding job cost about 1.4 cents. We cannot know Carbon's cost per job until we test it. Our price comparison shows how big those differences can be.

What does Google say it can do?

Google shares a big table with test scores. It compares Carbon with Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. These are Google's numbers on tests Google picked:

• Artificial Analysis Intelligence Index v5.0: 59, the top score. This index mixes many tests and is run by an outside company. Its own page agreed when we checked: Carbon (High) is first with 59, then Claude Opus 5.5 (Max) with 58.
• Terminal-Bench 4.0, a test of coding jobs in a terminal, the window where you type commands: 65.2%, against 59.6% for Opus 5.5.
• SEC-bench Pro, a test of finding security holes: 88.2%.
• ExploitGym, a test of using security holes: 54.8%, against 42.4% for Astra.
• GraphWalks, a test of very long texts: 96.2%.

Google ran the security tests without its usual safety limits, "evaluated without guardrails" in its words. So these scores do not show how Carbon does with safety limits switched on.

On most rows, Carbon has the best score. But one test score does not tell you how a model does on your own work.

Is it the best model now?

Not on every test. It is worth reading Google's table to the end:

• FrontierSWE 2, a test of long programming jobs: GPT-6 Astra scores 65.5% and Claude Opus 5.5 62.3%. Carbon scores 61.2%.
• OSWorld-2.1, a test of using a computer by clicking and typing, on the part that runs offline: Claude Opus 5.5 scores 87.2%. Carbon scores 81.5%.

So Carbon leads on the mixed index and on security tests. On some coding and computer jobs, older rivals still score higher in Google's own table.

Safer than Argon?

Google spends a large part of the post on safety. A few claims stand out:

• Prompt injection: that is when hidden text on a web page or in a file tries to give the AI orders. On Gray Swan's test, 0.0% of the attacks worked on Carbon, against 0.7% on Argon.
• Cheating on tests: on Terminal-Bench 4.0, Google says Carbon tried to cheat in 0.4% of the jobs, and none of those tries worked.
• Doing too much or too little: Google says Carbon did 26% less "overreach", going beyond what the user asked. It also did 29% less "under-service", stopping too early.

All of these are Google's own measurements, mostly on its own tests. We could not repeat them.

Can developers find it yet?

Not yet. On 29 October 2026 we checked the places developers would normally look:

• The Gemini API model list did not mention Carbon.
• The Gemini API price list did not mention Carbon.
• The newest entry in the API update notes was from 27 October, about a new command-line tool. There was no entry for Carbon.

We also looked on OpenRouter, a service that gives access to many models. Carbon was not there either. So the prices above are the prices Google announced. We found no public Carbon listing on the developer pages we checked.

What can you use today?

Google says access will start with a small group of cyber defenders. We could not confirm who already has access, and Carbon was missing from the public developer lists we checked. These models you can use today, with our own results on the same three secret coding jobs:

• You want Google's best model you can buy: Gemini 4 Argon. It passed every check in our test. Its price doubles on 1 December.
• You want Carbon: wait for Google's next step. We will test it as soon as we can and add a dated update here.

Typical time is the middle value of nine runs: each job three times. Our jobs are small and clear. They do not show how a model does on a long project.

How we checked

On 29 October 2026 we read Google's announcement of 28 October, including the note with the later price and the benchmark table. The same day we checked the Gemini API model list, price list and update notes, OpenRouter's model list, and the Artificial Analysis index. We also read the price pages of OpenAI and Anthropic for the comparison.

We could not use Carbon ourselves, because it was not open to us. All claims about how well it works are Google's, except the index score, which we checked on Artificial Analysis's own page. The results in the last table come from our earlier runs of our secret coding test, between 6 and 17 October 2026, explained on our methods page. The screenshots are our own captures of Google's public pages, only cropped. The charts are our own, and none of the images was made by AI.

This is a summary of public sources, not a test of Carbon. We will add a dated note when more people can use it.