Google prices Gemini 4 Argon at a fifth of the OpenAI and Anthropic flagship list price to buy back the frontier: Two dollars in, ten dollars out, per million tokens. Through November 30. On December 1, the rate doubles to $4 and $20, and Gemini 4 Argon joins Claude Opus 5.5 and Meta's Muse Wonder at exactly their list price.
That is a seven-week window, not a repricing. And it is, per The Inference, reportedly the only frontier discount in the market currently funded by a public income statement.
Last week we mapped the split between AI labs that answer to capital markets and labs that answer to private term sheets. This week the pricing lever itself became the story: Google is not selling Argon cheap so much as buying workload migration, and the receipt will land in Alphabet's next quarterly filing.
Google is not selling Argon cheap so much as buying workload migration, and the receipt will land in Alphabet's next quarterly filing.
Start with what the discount actually squeezes. OpenAI already sells GPT-6.1 Sol and Anthropic sells Claude Sonnet 5.5 at $2 and $10. So Google's introductory rate does not undercut the market's mid-tier — it opens level with it. What the headline number attacks is the $10/$50 flagship tier: GPT-6 Astra and Claude Fable 5.1. Every token Google serves at $2 is a token OpenAI has to defend at $10.
The awkward part is that the discount lands on a model that is not the capability leader. On Google's own comparison chart, Astra still beats Argon on FrontierSWE v2, 65.5% to 55.0%, and Argon trails on Terminal Bench 4. Artificial Analysis scores Argon at 53, tied with Astra and Fable 5.1, below Opus 5.5 at 58 and Muse Wonder Full at 56. Sam Witteveen of Red Dragon AI called it "a really, really, really good model" — then added that he does not think it is the best one out there.
So the offer is not "we are smarter." It is "we are cheaper, and we are available now." That is a customer-acquisition subsidy with an expiry date, and the date is the point. Workloads that migrate in October do not migrate back in December because a list price moved. The seven weeks are, by the article's own characterization, a switching-cost trap dressed as a sale.
Now follow the money, because this is where Google's position differs structurally from its rivals'. Alphabet earned $40.8 billion of operating income on $119.8 billion of revenue in Q2 2026 — a 34% margin — and spent $44.9 billion on capital expenditure. That was the first quarter in two years that capex outran operating cash flow of $39.1 billion. Meta spent $30.1 billion in the same quarter.
The labs charging $10/$50 are private. OpenAI carries a $1.18 trillion valuation, Anthropic $350 billion, and neither has to mark a discount against a quarterly print. Google does. The subsidy is visible, auditable, and — for now — affordable only because search and cloud still throw off that 34% margin.
This is what the capital constraint looks like when it stops being an analyst theme and becomes a line item. The company with the balance sheet is using it to buy share from companies without one.
There is a counterargument worth taking seriously: that Google is simply pricing at cost and letting volume do the work. The evidence for that is thin but real. Argon High's average cost per task is $1.99, about 39% below Astra. That is the only per-task cost figure any vendor has put on the record, and it is the first partial answer to the question of whether vendors would publish real cache-hit ratios and cost per solved agent trajectory.
Partial, because Google still has not published token consumption per task. And that omission is where the pricing story turns from clever to genuinely unresolved.
List prices are per million tokens. Work is per completed task. Those two numbers diverge precisely where the frontier is heading: long-horizon agent loops that read, call tools, revise, and check results across many steps. Google's own launch framing is explicit about this — Argon's core upgrade is sustained execution on long multi-step work, not single-shot answers.
A model that solves a task in one pass at $10 per million tokens can be cheaper than a model that wanders for forty steps at $2. The evidence pack here is candid about the risk: if the model is not smart enough and burns tokens going back and forth, the savings burn right back. Futurum's analysts flagged the same thing in plainer language — Argon's weaker points include the number of tokens it consumes.
Which means the scoreboard everyone is reading may be the wrong one. A ratio of list prices tells you what a vendor charges per unit of input. It does not tell you what a customer pays per unit of finished work. Argon's $1.99 figure is the closest anyone has come to publishing the metric that matters, and it arrives without the denominator that would make it comparable.
Consider what Google is actually claiming internally. Argon worked autonomously for 11 hours on Google's Agent Development Kit, a 205,000-line Python project, and returned six optimization proposals spanning 32 files after running all 16,892 open-source tests; two were merged. It helped reorganize roughly 6 million lines of Chrome feature code inside a 40-million-line Chromium codebase, work Google says was expected to take months and finished in weeks. It optimized Gemma 4 E2B inference from about 37,900 to over 108,000 tokens per second.
Those are task-level results, and they are the right kind of evidence. They are also company-reported and unreplicated. The gap between "Argon did this once inside Google" and "Argon will do this for you at $2 per million tokens" is the entire distance between a pricing stunt and a durable cost advantage.
Which brings the question back to the buyer. If you are running agent loops at scale, the number that decides your architecture is not $2 or $10 or $50. It is dollars per solved trajectory, and no vendor has published it — not Google, not OpenAI, not Anthropic. Argon's $1.99 per task is a first crack in that wall, and it came with a promotional clock attached.
So read the seven weeks for what they are. Google is buying migration at a moment when its capex has outrun its cash flow, against private rivals who cannot be audited the same way, on a model that trails a rival on the benchmarks Google chose to publish. The discount is real. The price level is temporary.
The unresolved question is whether any of it survives contact with an agent loop. If Argon's token appetite on long-horizon tasks runs hot, December's $4/$20 will feel less like a price increase and more like a correction to a number nobody had measured properly in the first place. The vendor that publishes cost per solved trajectory — with the token counts attached — will set the next scoreboard. Until then, everyone is arguing about the wrong number.