Google opened its most capable artificial-intelligence model to developers, enterprises and consumers on Tuesday, and put the detail that matters most to buyers in a footnote.
Gemini 4 Argon is now available worldwide through the Gemini API, the company said, and through Google Antigravity, the Gemini app and AI Mode in Search for paid Google AI subscribers. Google described it as its strongest model for what the company calls deep work: complex, real-world tasks in coding, finance, legal, cybersecurity and science.
The benchmark claims reach wider than a single category. Google said Argon outperforms other frontier models on multi-step tasks and ranks first on the Vals Index, which scores model performance in finance, coding, legal and tax work, weighting each sector by its contribution to U.S. gross domestic product.
Then comes the price. For an introductory period, Argon costs $2 per million input tokens and $10 per million output tokens. Cached input tokens are discounted 95%, to 10 cents per million. On Dec. 1, the rates become $4 and $20.
Output tokens in that schedule include the reasoning the model generates before it answers, which is where agentic products spend most of their money. A coding agent that reads a large repository, plans a change and checks its own work can consume millions of tokens to produce a few hundred lines of code. Google said input, thinking and output share one window of about a million tokens.
The discount on cached input is the quieter lever. Systems that keep a long prompt or a codebase pinned in context pay a tenth of the input rate to reuse it, which changes the economics of repeated calls against the same documents. Legal research, tax review and large-scale code refactoring all fit that shape, and all are categories where Google is now claiming an advantage.
Argon is not on Google’s free tier or its AI Pro plan. Consumers reach it through the Gemini app and AI Mode in Search on subscription tiers, while enterprises reach it through the API. That split says where the company expects the volume to come from.
Analysts said the introductory price is a customer-acquisition tool as much as a technical launch. Frontier models are broadly comparable on headline benchmarks, so the decision for a large buyer often comes down to cost per completed task, data residency and whether the vendor offers a path off today’s rate. A two-month window gives Google a way to win pilots without committing to a price it may not want to defend later.
The window also creates a decision problem for the buyers it is courting. Teams that build a production workflow on Argon at $2 and $10 are betting that Google extends the discount, or that Argon will still be the best model available when the bill doubles. Neither assumption is under their control.
Google has not said whether the introductory period can be extended past Nov. 30, or whether the standard rates take effect in every market on the same day. The company’s published case for the model rests on benchmark position and price rather than on long-term contracts, and benchmark leadership in this market has rarely survived a full year.
One reading of the announcement is that it is a land grab. The company that gets developers to write their first agent against its API also captures the retry loop, the tooling and the monitoring layer built around it. Switching costs rise with every integration. Offering the strongest model at half price for two months is an efficient way to buy that position, and Google can afford the promotion because it sells the compute as well as the model.
A second reading is that the standard price is the real story and the discount is packaging. If token prices across the industry keep falling, a $4 and $20 list looks less like a premium and more like a reference point competitors will undershoot. Enterprises that sign annual commitments at introductory rates would then be paying above market by spring.
The financial disclosure is thinner than the pricing sheet. Google did not break out revenue expectations for Argon, the cost of serving it, or how much of its existing fleet has been converted to run it. Alphabet has been under pressure to show that its spending on data centers converts into cloud revenue, and a flagship model reserved for paying customers is one way to make that argument.
Google declined to say how many customers are running Argon in production today. A developer who builds on the API put the trade-off plainly, asking not to be named to speak about a vendor his employer depends on: the model is the best available for his team’s task this month, and the price is the only part of the arrangement he can be certain will change.
For a chief technology officer, the question is narrow. Argon is the strongest model on the index Google cites most, and for now it is the cheapest way to run a frontier system at scale. Both halves of that combination are scheduled to move on the same day, and only one of them is a choice.
