From a vetted list to general sale
Google has made Gemini 4 Argon generally available, thirteen days after releasing it only to cyber defenders approved through its Fairwind Program. The developer announcement puts the model in the paid tier of the Gemini API through Google AI Studio and Android Studio, and in Google Antigravity — which Google describes as an agent-first development platform — for Google AI Ultra subscribers.
Google calls Argon “our frontier model for complex workflows” and aims it at software engineering, enterprise knowledge work and cybersecurity defence. The post describes the target as “long-horizon workflows that require deep reasoning and structural precision”, naming code migrations and frontier research as examples.
The launch price has an expiry date
The number that matters is on Google’s pricing page, not in the blog post. On the paid tier, Argon costs $2.00 per million input tokens and $10.00 per million output tokens — output including thinking tokens — through 30 November 2026. From 1 December the listed rates are $4.00 and $20.00. Cached input follows the same pattern, going from $0.10 to $0.20 per million, and only implicit caching is supported: explicit caching is not available on this model.
Google does not use the word “introductory” on that page. It simply lists two sets of prices with a date between them. Anyone budgeting a migration that runs past November is budgeting at the higher number.
What “one million tokens” now means
At the 30 September launch, Google said it was “significantly expanding the model’s output token limit to an industry-leading 1M tokens”, up from 64,000. The documentation now describes the same figure differently. The pricing page calls the model capable of “long-decode generation up to 1M tokens”, and the model documentation refers to “a shared 1M context window” together with “long decodes up to 1M tokens”.
A shared window and a separate output ceiling are not the same thing: in the first case a long prompt eats into what the model can still write. Google has not published a figure for the longest single uninterrupted response.
Where it sits when someone else is counting
Google’s own benchmark claims from the launch post are specific. It reported 77.9% on DeepSWE v1.1, which it called a new state of the art; 51.3% on Zapier’s AutomationBench, ranked first; 91.7% on the long-video benchmark LVBench; and 68% on CWE-bench v1, tying for first place. Those are vendor-reported figures.
Artificial Analysis, which runs its own evaluations, scores Gemini 4 Argon at 53 on its Intelligence Index and ranks it eighth of 216 entries. On that leaderboard Claude Opus 5.5 leads at 58, Claude Sonnet 5.5 and Meta’s Muse Wonder Full follow at 56, and Argon ties on 53 with GPT-6 Astra and Claude Fable 5.1. Independent indices and vendor benchmarks measure different things, and in this case they disagree about where the model stands.
What to watch
Two dates. The first is 1 December, when the listed price doubles and the economics of any long-running agentic workload built on Argon change with it. The second is whatever Google says about the output limit, which is currently described one way in a launch post and another way in the documentation developers actually bill against.
