Gemini 4 Carbon: $2 per million tokens, but the door is still shut

On Wednesday 28 October 2026, Google unveiled Gemini 4 Carbon on its official blog, calling it its most powerful model yet. It arrives just two weeks after Gemini 4 Argon. At this rate, the Gemini 4 family is working through the periodic table like a student cramming the night before a chemistry exam.

What Google puts in the shop window

Every figure below comes from Google, which also happens to be the company selling the model. Keep that in mind. Carbon reportedly takes first place in version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. In cybersecurity, the core of the announcement, it scores 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploiting them) and 71.0% on CWE-bench (fixing them). Google adds that these external tests were run without safeguards. In other words, the model being tested had its guardrails removed, so it isn’t the version you’ll actually get to use.

On its own tests, Google says Carbon is more than twice as successful as Argon at generating secure code. According to the Chrome team, Carbon writes patches more than ten times better than those of a leading frontier model from another lab. Wiz, Google’s cloud security platform, ranks it first on its web penetration tests. For web development, Carbon gains 89 Elo points on WebDev Arena and ranks fourth, at 1768.

A model that won’t let it go

The other promise is about long-running tasks. According to Google, its teams have watched Carbon work for days on multi-step engineering problems. One example: tracing a subtle bug back to its source and then fixing it, in a large codebase or in low-level code, without giving up at a dead end. It’s the colleague who won’t go home until the test goes green, and who sends you the compute bill the next morning.

Google also says that in real-world use with its own teams, it saw lower rates of misalignment and fabrication. Put simply, the model apparently makes things up less often. The announcement doesn’t include a single figure to back that up.

And against GPT-6.1?

On this question, the announcement is silent. Google’s post doesn’t mention OpenAI or its GPT-6.1 line-up, and it publishes no comparison with any named competitor. The only external comparison is the Chrome team’s, which pits Carbon against a model from another lab without naming it. You can probably guess which one, but you can’t check, and neither can we. For now, any head-to-head figures between Carbon and GPT-6.1 would be guesswork, so we’ll leave that to the people reading tea leaves.

A price on the tag, a lock on the door

The pricing, at least, is clear. Carbon will launch at the same introductory price as Argon: $2 per million input tokens, $10 per million output tokens, and a 95% discount on cached input tokens, which works out to 10 cents per million. Google calls this a launch price, and nothing says it will last.

Paying it is another matter. Google has chosen a gradual rollout and is taking part in the US government’s voluntary process for testing models before release. Starting this week, Carbon is available to some cyber-defence specialists. Access will then widen step by step to other trusted partners through the Fairwind programme. Developers, businesses and the general public come after that, “as soon as possible”, with no date given.

So you now know the per-token price of a model you can’t call yet. It’s like a restaurant that tapes its menu to the window before opening: you can read every price, but nobody will take your order.