Google’s September 30 post announcing Gemini 4 Argon was, by the standards of these things, unusually specific. The model launches at $2 per million input tokens and $10 per million output tokens, with cached input discounted 95 percent, to $0.10 per million. Then, “after the introductory period expires,” the same tokens cost $4 and $20.

That second sentence is the news. Almost nobody read it.

The Price War Has an Expiration Date Written Into It

VentureBeat framed the launch correctly: Google is pairing benchmark claims with aggressive pricing. Argon’s introductory rate is one-fifth of OpenAI’s listed $10/$50 for GPT-6 Astra, and half of what Anthropic lists for Claude Opus 5.5. Everyone did the ratio. Nobody did the calendar.

A company that expects its product to commoditize does not announce a 100 percent price increase in the document that introduces it. It announces the increase later, quietly, after switching costs have set. Google announced it on day one, in public, in the same post as the benchmark table — which means either that Google believes frontier capacity will still be scarce when the introductory period lapses, or that it is willing to bet its developer base on that belief. Both readings are bad news for the commoditization thesis that has been the load-bearing assumption of three years of enterprise AI planning.

The 95 percent cached-input discount deserves the same reading. Ten cents per million cached tokens is not generosity; it is an architecture subsidy. Cached input is cheap for Google because nothing is being recomputed — you are re-sending the same context, which is precisely what long system prompts, document corpora, and agent memory do. Price that at a tenth of the going rate and you get a generation of stacks shaped around cached context. The more your product is shaped around it, the more expensive it is to leave.

“We rebuilt our retrieval layer around cached input in August,” one staff engineer at a payments company wrote in a Slack DM the day Argon shipped. “The new pricing page doubles our bill in six months. My CFO finds out about that in Q1, and it will be my problem.”

A Frontier Model That Ranks 32nd

The independent tracker BenchLM scores Argon 64.6 out of 100 and places it 32nd of the 211 models it follows. That is not a scandal and not proof of anything — different evaluators weight different things. It is a reminder that “retakes the benchmark lead” is a genre with a house style, and the house style is to pick the configuration.

Google’s own launch card is more candid than the discourse around it. The fine print states that OSWorld 2.0 is a partial score from the offline subset, that Terminal-Bench-Science 0.1 uses a sixfold verifier timeout, and that launch scores are “maximized over three runs.” Max-over-three-runs is a choice that raises numbers. Google disclosed the choice. That is not deception; that is convention — and convention is doing the work that deception would otherwise have to do.

Then there is the figure nobody quoted: Harvey’s legal agent benchmark, at 19.6 percent. It sits in the same post as “leading performance across other domain specific evaluations” and AutomationBench’s 51.3 percent, where Google ranks first on Zapier’s end-to-end business benchmark. All three can be true. Only one traveled, because “leads on legal work” is not the sentence the launch was for.

‘Rolling Out Soon’ Is the Disclosure

Argon goes first to Google AI Ultra subscribers and paid API customers. Rolling out soon. That is a strange way to launch a frontier model in a business that spent two years giving frontier models away to anyone with a login, and it reads like a capacity decision: the model is scarce, and Google is behaving as though scarcity is something to monetize rather than something to apologize for.

Note also the spec the launch card goes out of its way to clarify: a 1 million-token output limit, not an input context window. The distinction sounds technical. It is not. It is the difference between a model you can feed your archive and a model that can write you a book — and the fact that Google’s own documentation has to correct the reading tells you what the leaks had been promising for months.

Who Has to Rewrite Their Forecast

Three groups. First, every enterprise that built a twelve-month inference budget on a declining unit-cost curve. The curve just got a step in it, and the step is documented. Second, the commoditization chorus — the people who have spent two years explaining that intelligence is a commodity and model margins are a rounding error. Google’s price list is now the best counterargument available to them. Third, and least comfortably, the people who cheered the price war on the theory that competition drives prices down.

It did. And then, on a schedule printed in the launch post, it stopped. That is not a regulatory failure or a market failure — it is what a market with three credible frontier suppliers and one shared cost structure looks like when it works. Nothing was hidden. Nothing was illegal. Your tokens still get more expensive in six months.

The people who told you cheap intelligence was a permanent condition were not lying to you. They just were not reading the footnotes.

Sources