Google released Gemini 4 Argon at $2 and $10 per million tokens

A team that was waiting on Google's next frontier model to budget its next agent build has the two numbers it needed. Google made Gemini 4 Argon generally available, with an introductory price of $2 per million input tokens and $10 per million output tokens on the Gemini API, cached input at 95 percent off, and reach through the Gemini app, AI Studio, Vertex AI, and the Gemini CLI. The release follows a two-week preview in which the model was open only to a set of cybersecurity teams through Google's Fairwind Program.

The model has a 1 million token window combined across input and output, which Google calls a unified limit. Access runs across the paid API, Google AI Ultra inside Antigravity, the Gemini app for Ultra and Pro subscribers, and AI Mode in Search for the same subscribers. Google said the price will stay at the introductory level before moving to a standard rate it did not name.

What Google says it tested

Google published a set of its own benchmark results beside the launch. On the Vals Index, which tests coding, finance, legal, and tax work, Google said Argon ranks first. On the Embodied Reasoning benchmark for robotics, the company reported 87 percent. On RareBench, a test for rare disease diagnosis questions, it reported 57 percent. Google said Argon scored under 1 percent on its own bio red-teaming tests, which measure whether the model will help with harmful biology work. Every one of these numbers is Google's own, run and reported by Google on its own test suite.

The company also described a safety process that includes prompt injection defences, misalignment monitors that watch the chain of thought for signs the model is stepping outside the user's intent, and findings that are deliberately not fed back into training so the model is not taught to hide its reasoning. These controls were in the earlier September post as well.

How this differs from the September release

Google first announced Argon on 30 September and released it to a set of cyber defenders through its Fairwind Program ahead of paid API customers, which the company said would get access when the safeguards work was done. That release carried the same prices and the same 1 million token output limit as today's launch, but the window today is phrased as combined input and output, not output only.

The author of the launch post is Koray Kavukcuoglu, senior vice president of Google DeepMind and chief AI architect at Google. He described Argon as built for "deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science," and said the model is Google's offer for teams that want one model for both long reasoning and tool use rather than one for each.

For a team that was already planning around Gemini 3.8 Flash or Pro, the question today is a plain one. The price per task of running an agent on Argon can be worked out from the input and output ratios the agent already uses on its current model, and the 1 million token window answers whether a particular long-document task that was split across calls can now go in one.

Source

Introducing Gemini 4 Argon, by Koray Kavukcuoglu on Google's blog.