Gemini 4 Argon: Availability, Pricing and Output Limit

What Google announced

Google announced Gemini 4 Argon on September 30, 2026, calling it a new frontier model and the start of its next era of models. Google aims it at long, multi-step professional work: real-world software engineering, enterprise knowledge work in legal and finance, and cybersecurity defence. It is the first model carrying the Gemini 4 name, and it arrives as a limited release rather than a general launch.

Who can use it, and when

This is the most important fact on the page. Gemini 4 Argon is not generally available. Google is rolling it out first to a set of trusted cyber defenders through what it calls the Fairwind Program, and says it will gather feedback from those early testers and refine its guardrails before releasing it to developers, enterprises and consumers, starting with paid API customers and Google AI Ultra subscribers. Google has not given a date for that wider release. Google also describes taking part in the U.S. government's voluntary process for pre-release model access. No public API model ID has been published yet. If you cannot see Argon in your Google account or API console, that is expected: for almost everyone, this model is announced, not released.

Pricing, once it opens up

Google has already set the price. Argon will launch at an introductory $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input price. After the introductory period ends, the price becomes $4 per million input tokens and $20 per million output. Google has not said how long the introductory period lasts. Because the model cannot be bought yet, we show these as announced prices and leave Argon out of our price charts and value comparisons until it can actually be used at them.

The 1M-token output limit

The one hard specification Google gives is the output limit: 1 million tokens, up from 64,000. That is unusual. Most current models allow somewhere between tens of thousands and a few hundred thousand tokens of output per request. Google's reasoning is that a model working through a long, hard problem needs room to reason and produce hundreds of thousands of tokens in a single run. Note what this is not: it is the output limit, not the context window. Google's announcement does not state Argon's context window, so we show it as not stated rather than assume it matches the output figure.

Google's published results, and how to read them

Google reports 77.9% on DeepSWE v1.1, 91.7% on LVBench, 68% on CWE-bench v1 and 51.3% on AutomationBench, and says Argon leads the Vals Index without giving a figure. These are Google's own results under its own settings, from a model almost nobody outside Google and the Fairwind Program has been able to test. We will add independent measurements, including an LMArena Elo, once the model is broadly available and those boards have rated it.

What to do with this today

For most teams, nothing yet beyond planning. The announced pricing is useful for budgeting ahead of general availability, and the output limit is worth noting if your work involves generating very long documents or long agent runs. Everything else, from real-world latency to how it handles your own tasks, has to wait until access opens. We will update this page when Google publishes an API model ID, a context window and a general availability date.

Official Resources