Google released Gemini 4 Argon on September 30, 2026, starting with a limited rollout to cyber defenders rather than a general launch. According to Google’s announcement, the model raises the output token limit to 1 million tokens, up from 64K, and comes with Google-reported scores on coding, workflow automation, long-video and vulnerability benchmarks.
On this page
Who gets access first
Google says Gemini 4 Argon is “rolling out to a set of trusted cyber defenders through our Fairwind Program.” The company also says it is “actively engaged in the U.S. government’s voluntary process for pre-release model access.”
Wider availability is planned for “developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers,” which Google says will come “as soon as possible.” Until then, the model is limited to the Fairwind Program group.
A 1 million token output limit
The largest specification change in the announcement is output length. Gemini 4 Argon can return up to 1 million tokens in a response, compared with the 64K limit Google cites as the previous ceiling. For developers, that means much longer single responses, such as large code changes or long documents, without splitting the work across many calls.
Benchmark results
All scores below were published by Google in its announcement.
| Benchmark | Gemini 4 Argon | Google’s note |
|---|---|---|
| DeepSWE v1.1 | 77.9% | — |
| AutomationBench (Zapier) | 51.3% | Ranked #1 |
| LVBench (long video) | 91.7% | — |
| CWE-bench v1 | 68% | Tied for first |
Google also reports leading results on the Vals Index and on Gray Swan’s indirect prompt injection evaluation. Lab-published results for this and other releases are collected on our benchmark leaderboard.
Safety measures
Google says the model is designed to “refuse harmful requests while preserving legitimate, dual-use scientific research.” The announcement lists “misalignment mitigations,” “prompt injection robustness” and the use of “sandboxed environments” as part of the safety work around agent use.
The rollout order follows the same line: the first users are security teams admitted through the Fairwind Program, ahead of developers and consumers.
What it changes
For now, Gemini 4 Argon is not a model most developers can call. Teams outside the Fairwind Program will have to wait for the paid API and Google AI Ultra rollout Google describes. When that happens, the 1 million token output limit is the change most likely to affect how applications are built on Gemini, since it removes the need to chunk long generations.
Gemini 4 Argon is listed on our AI model release timeline, and service availability for Gemini is tracked on the Gemini status page.
