Google opens Gemini 4 Argon to paying customers

The announcement is dated Oct. 13, 2026. The byline is Koray Kavukcuoglu. Thursday’s File, on Oct. 1, reported the Sept. 30 post: first release only to trusted cyber defenders, with broader access planned later.

Google now says Argon is globally available for developers and enterprises through the programming interface in AI Studio and the Gemini platform. Google AI Ultra subscribers can use it in Antigravity. Pro and Ultra subscribers can use it in the Gemini app and in AI Mode in Search. A second post says the paid tier is also in Android Studio.

The introductory price is $2 per million input tokens and $10 per million output tokens. Cached input, meaning a repeated chunk the system has already stored, is 95 percent off the input price. A footnote says that after Nov. 30, 2026, the price is $4 per million input tokens and $20 per million output tokens, and that this later price applies to the Argon model intended for public access.

Google says the model can take in and write out up to 1 million tokens in one job. The model card, published October 2026, says Argon is not a modification or a fine-tune of a prior model. It takes text, images, audio, and video. A fine-tune is a smaller training pass on top of a model that already exists.

A link on the announcement opens a Google Cloud console with the model parameter gemini-4-argon. The developer guide page this desk opened did not print a separate model-ID line in the text that came back.

On cyber defense, Google said it is beginning to add hundreds of additional Fairwind partners. Defenders can still apply for the version without cyber guardrails. The Sept. 30 limit was the first release. This is the expansion.

The safety post, by Anca Dragan and Raluca Ada Popa, says an outside biology test could unlock less than 1 percent of the bio tasks they score. On a coding test where the model had no reason to hold back, shortcut attempts fell to 2 percent. On a prompt-injection test from Gray Swan, attack success across 15 tries was 0.7 percent. Prompt injection means hidden instructions in a web page or a file that try to hijack the model. Those figures are Google’s. The same post says the model solved all 12 of Google’s simulated end-to-end cyberattack scenarios, up from five in the previous generation. That, too, is Google’s test.

The safety post and the announcement both point to a model card and a longer safety report. This desk read the model card. The safety report file was large. The claims above are from the posts, not from a line-by-line reading of that report.

Google also describes uses inside the company, from code cleanups to science. Those are the company’s examples. This story does not re-score them.