Gemini 4 Carbon leads one long-context test. It is not on sale to everyone

Google posted Gemini 4 Carbon on 28 October and called it the strongest Gemini yet, two weeks after Gemini 4 Argon. The post, from Koray Kavukcuoglu, says the gains are in coding, agents, science, and cybersecurity. The part of the published table that is not close is long context. On GraphWalks, a walk across a graph inside a prompt of 256,000 to a million tokens, Carbon scores 96.2 percent. Argon is at 84.2. GPT-6 Astra is at 71.8. Claude Opus 5.5 is at 66.8. That is Google’s table, with a methodology page linked under it. A high score on one retrieval test is not a finished agent.

The same table is less dramatic everywhere else, and it does not give Carbon every row. On Artificial Analysis Intelligence Index v5.0 the scores are 59 for Carbon, 58 for Opus 5.5 and 57 for Astra. On FrontierSWE 2, Astra is ahead, 65.5 to 61.2. On OSWorld, Opus is ahead, 87.2 to 81.5. Where Google highlights cybersecurity, the table shows SEC-bench Pro at 88.2 percent, ExploitGym at 54.8 and CWE-bench at 71.0. The prose also says secure-code generation more than doubled against Argon on an internal test, and that high-severity misalignment alerts in live coding fell by almost 80 percent. Those last two are internal. They are not rows a reader can rerun.

Who can call it is the narrower fact. Google says it is in the US government’s voluntary pre-release testing. Starting this week, Carbon goes to a subset of cyber defenders, then to other trusted partners through a programme it calls Fairwind. Developers, businesses and ordinary users come later, after the guardrails have been changed in response to that feedback. “As soon as possible” is not a date. A model that is first in a defender’s hands is not yet a model a robotics team can put on a robot.

The introductory price is the same as Argon’s: $2 per million input tokens and $10 per million output tokens, with cached input at 95 percent off. Google’s footnote says that after the introductory period the price becomes $4 and $20. It does not say when the period ends. Until the model is on the public API, the table is a claim and the price is a plan.