Google opened its flagship model Gemini 4 Argon to every developer and enterprise customer on October 13, ending a limited release that had kept the model behind a security program since late September.
The gate had been narrow. For weeks Argon ran in the hands of trusted cyber defenders under Google’s Fairwind program, who received it without the usual cyber guardrails so they could hunt software vulnerabilities with it. On Tuesday the model moved onto the Gemini API, AI Studio and Antigravity, the company’s agentic development environment.
The specifications are the sales pitch. Argon takes a context window of one million tokens on both input and output, and the output limit is the more consequential of the two claims, since most competing models accept long prompts but write short answers. Cached input tokens are priced at a 95% discount to standard input.
The output limit is the part developers will notice first. An agent working through a multi-step task has to hold the plan and the results at the same time, and a million-token ceiling on the response lets a single call return a working file instead of a fragment of one. That changes the shape of an application more than the price does, because it removes the orchestration layer that used to split long jobs across calls.
The prices are introductory: $2 per million input tokens and $10 per million output tokens. Google did not say how long the introductory period lasts, and that omission matters more than the number itself.
The internal figures Google published alongside the launch are the more interesting disclosure. More than 100,000 employees now use the model. Across four weeks, engineering teams working with agent-written code increased their commit volume by 35%, and the share of commits that had to be rolled back fell by 30%.
Read those together and the claim is not that the model writes more code. It is that it writes code that survives review. Rollback rate is the measurement that costs a company money; anyone can generate volume, and the second statistic is the one a chief technology officer would ask for.
Google is not selling Argon into a quiet market. Meta priced its first frontier model, Muse Wonder, at $4 per million input tokens and $20 per million output when it launched a day earlier, and the strongest Claude and OpenAI models sit above that. Argon’s entry price undercuts the field by half on input, which is a statement about Google’s cost of serving, not about its generosity.
A model company that designs its own accelerators can price an inference request closer to its marginal cost than a rival buying the same chips on the open market. The strategic question is whether the introductory rate reflects a durable cost advantage or a promotional window built to pull workloads onto Google’s stack before competitors answer with cuts of their own.
For buyers, the decision is harder than a price comparison. Migration costs on a frontier model are real: prompts get retuned, evaluations get rebuilt, and agent scaffolding is written against a specific tool-calling behavior. An introductory price invites that investment and then expires. Developers who tested Argon during the limited release advised treating it as strategically important now while holding off on a full production migration until the API surface and the model identifiers settle.
The context window introduces a second pricing problem. A million-token input makes a single request expensive in absolute terms even at a low per-token rate, and workloads that push a codebase or a document set into the prompt generate bills that scale with how much context the agent retrieves rather than with how many questions it answers. The 95% caching discount is where the economics actually get decided. Teams that structure requests so the stable part of the context is reused pay a fraction of what a naive implementation costs.
That is a design constraint dressed as a discount. It rewards customers who build caching into the architecture from the start and penalizes those who bolt it on later.
Analysts said the launch keeps Google in the position it has held for two years: first to publish aggressive prices, with the expectation that rivals match within a quarter. In that pattern the release matters less than the price sheet attached to it.
Analysts said the three surfaces matter for a second reason: they let Google watch how the same workload behaves in a prototype, inside an agent loop and in production before it commits to a permanent rate. A price set after seeing that data is a different decision from a price set to win an announcement.
Google’s other bet is on defensible work. Argon was trained for cybersecurity defense and can find, validate and patch vulnerabilities on its own, which is why the first release went to defenders. An enterprise buying the model for security work is buying a capability a general-purpose chatbot does not have, and that is a harder thing to commoditize than a token price.
What the launch materials do not address is how the company expects the economics to look once the promotional period ends. Google said only that the discount is limited. A price with an unstated end date is a negotiating position, and every procurement team that signs a multi-year commitment this quarter is making an assumption about what happens when the clock runs out.
The model is available now. How long the rate stays available is the open question.
