Gemini 4 Argon is now available to paying developers and subscribers, at an introductory price until November 30

You can now call Gemini 4 Argon, Google's newest frontier model, through the paid tier of the Gemini API in AI Studio and Android Studio. When Google announced it on September 30, 2026, only trusted cyber defenders in the Fairwind Program could use it. The rollout that started on October 13 is global and also covers enterprises and Google AI Pro and Ultra subscribers.

Who can use it, and where

Developers reach the model under the ID gemini-4-argon. Enterprises get it through the API, in Gemini for business, and in Antigravity, Google's agent-first development platform. Google AI Ultra subscribers can also try it in AI Studio and Antigravity.

Everyone else gets Argon in the Gemini app and in AI Mode in Search, but only with a Google AI Pro or Ultra subscription. The announcement does not mention the free tier of the API or free Gemini accounts.

Two groups get special terms. Google says it is extending the Fairwind Program to hundreds of additional partners, who receive the model without its cyber guardrails. Google is also offering 10,000 scientists six months of Antigravity's highest usage tier at no charge, subject to eligibility, capacity, and selection.

What it costs, and until when

The introductory price is $2 per million input tokens and $10 per million output tokens. Cached input tokens, the parts of a prompt the system has already stored, cost 95% less, which by our calculation is $0.10 per million. A footnote in the announcement says the introductory period ends on November 30, 2026.

After that date, the price is $4 per million input tokens and $20 per million output tokens, twice the introductory rate on both sides. This suggests that a project you start on Argon now should be budgeted at the higher rate, because the introductory price lasts less than two months.

What changes for code that calls the API

The context window, meaning the total amount of text the model handles in one request, is 1,048,576 tokens. According to Google's developer guide, input, thinking, and output tokens all share that space. A single request still returns at most 65,536 tokens when you stream the response, or 32,768 tokens for a non-streaming HTTP request.

Longer answers arrive in parts, up to 1,048,576 tokens in total. You set max_output_tokens above 65,536, and when a response ends with the finish reason CONTINUATION, the API returns a continuation_token that requests the next part. The Python library google-genai follows these tokens automatically from version 2.30.1, and @google/genai for JavaScript does so from version 2.29.1. Without the SDK, your own loop has to send the token back.

Argon has three thinking levels, LOW, MEDIUM, and HIGH, and the default for gemini-4-argon is MEDIUM. Caching is implicit only: prompts that share a prefix of 6,144 tokens or more are cached automatically, and explicit context caching is not supported. This suggests that applications which create caches explicitly will need a different approach for Argon.

What Google claims about the model

The benchmark results come from Google and have not been checked independently. Google says Argon ranks first on the Vals Index, which measures work in finance, coding, legal, and tax tasks. On ERQA, a test of spatial and physical reasoning, Google reports a score of 87%.

For agents, resistance to indirect prompt injection matters: a page or message the agent reads can hide instructions meant to take over its task. On Gray Swan's test, Google reports that attacks succeeded 0.7% of the time across 15 adaptive attempts.