Gemini 4 Argon

The defenders tried it first. Google announced Gemini 4 Argon on 30 September 2026. It first opened the model to cyber defenders in Google DeepMind's Fairwind Program (an early access programme set aside for the security teams of governments, critical infrastructure operators and big technology platforms, with more than 650 partners). Since 13 October, Google AI Pro and Ultra subscribers have been able to use the model in the Gemini app. The Gemini API's model page also lists gemini-4-argon as a stable release.

Now for the price. Per million tokens, input costs 2 dollars and output 10 dollars (thinking tokens included). These prices hold until 30 November; from 1 December both double. There is no free tier.

The real novelty Google is selling is the output limit. The previous model's limit was 64 thousand tokens. This one's is 1 million. The main uses on the list are long-haul coding, migrating large codebases, and finding software vulnerabilities and patching them on its own. There are figures too: 77.9 percent on DeepSWE v1.1, and a 0.7 percent attack success rate on the indirect prompt injection test. Mind you, every one of these is Google's own measurement, and there is no independent result yet. And to read a one-million-token answer from start to finish, you may need to take some time off work.