Google announced introductory pricing for Gemini 4 Argon of $2 07 per million input tokens and $10 per million output tokens on September 30, ahead of its launch. OpenAI already charges exactly that for GPT-6.1 Sol. Cached input costs $0.10 09 per million on both. Two frontier labs now share one sticker price.
The tie is the headline. The second number in Google's announcement is the story. Argon's rate is "introductory," and Google says it rises to $4 10 and $20 afterward. That is a 100% 25 increase on every token you buy.
Google has not said when. I do not know either, and nobody outside Google does. A builder whose margin only works at $2 33/$10 is running on a clock with no face. An app pushing 100 million 26 input and 20 million output tokens a month pays $400 today and $800 later.
Every Token Has a Shadow Price
The Shadow Price is the rate you plan your margins around. It usually sits above the rate on the page, and most vendors never write it down. Argon is unusual because Google printed it on launch day: $4 12/$20. The day's pricing facts fall into three buckets.
What Argon costs per million tokens today, what it costs later, and its output limit.
The Sticker is what the pricing page says today. Argon and Sol both list $2 08 input and $10 output. Cached input is $0.10 09 on each. TechCrunch reports that Sol costs one-fifth of GPT-6 Astra's standard token prices.
The Shadow is the rate you budget for. For Argon it is $4 11/$20, posted by Google with no date attached. OpenAI lists no scheduled increase for Sol.
The Spread is the gap between the sticker and what your workload actually pays. Output costs 5x 27 input on both models, so output-heavy agents mostly pay output prices. Cached input runs 95% 28 below fresh input.
Long jobs widen the spread. Beam reports Argon can now write up to 1 million 02 output tokens in one response, up from 64,000. A single maximum-length answer costs $10 29 in output at the sticker and $20 at the shadow.
So does the tie prove a price floor? I don't think so. It proves a reference price, one two labs matched because they read the market the same way. Nothing in it says prices can't fall further. What the evidence points to is a ladder with two rungs: $2 11/$10 as the sticker and $4/$20 as the shadow.
Who Pays When $10 Becomes $20
Why would Google print the higher price on launch day? A posted $4 11/$20 makes $2/$10 feel like a gift. It anchors developers before most of them can even buy. Proactive reports Argon is initially rolling out to trusted cyber defenders, with Google's internal teams also having access and paid API customers and Google AI Ultra subscribers next in line.
The benchmark table is the shiny distraction. Artificial Analysis scored Argon level with GPT-6 Astra at 60% 01 of the cost per task with discounted prices. Business Model Analyst read the footnote: that 60% 05 uses the introductory price.
Here is what the doubling does to a simple offer. Say you charge $30 30 a month per customer, and each customer's agent burns 1 million input and 1 million output tokens. At $2 31/$10 the tokens cost $12, leaving $18 of gross profit. At $4 32/$20 they cost $24, leaving $6.
Gross margin falls from 60% 33 to 20%. LTV:CAC is what a customer earns you over their life, divided by what it cost to win them. Say you spend $60 34 to land a customer who stays 10 months.
At the sticker, lifetime gross profit is $180 35 and the ratio is 3:1. At the shadow, lifetime gross profit drops to $60 36 and the ratio hits 1:1. Every new customer now returns exactly what you paid to find them, and nothing more.
If the business only works at $2 11/$10, you have built a reseller of Google's promotion with a login page on top. Google sets your margin. Google can also take it back, on a date it has not published. That is a distributor dressed up as a software company.
Where do the dollars sit? In that example, output is $10 37 of the $12 token bill, or 83%. Cut output by 20% 38 and you save $2 per customer. Cut input by 20% 39 and you save 40 cents.
The thing that actually produces your margin is boring: a log of every request and a test set of real tasks. The model is a supplier, and suppliers change prices. Keeping those records is repetitive work, and nobody will clap for it.
Whether Argon is the smarter model is a question outside pricing, so I'll stay on the bill. A wrong answer is still a token bill with nothing to sell. Beam reports Argon hallucinates 15% 15 of the time on Artificial Analysis's AA-Omniscience knowledge benchmark, against 54% for GPT-6.1 Sol and 51% for GPT-6 Astra at maximum effort. The Decoder found Argon matches Astra in independent tests but burns more tokens.
Do fewer wrong answers pay for those extra tokens on your tasks? Unclear. Only your own logs can settle it.
Why $2/$10 is a rental, not a price
Doubling the token bill cuts a 60% margin to 20%.
Take a $30 30-a-month offer where each customer uses 1 million input and 1 million output tokens. Gross profit falls from $18 31 to $6 32 at $4/$20. LTV:CAC drops from 3:1 to 1:1.
Google posted the higher rate but not the date.
Argon's introductory $10 10 output rate rises to $20 at some point Google has not named. Business Model Analyst notes that the last two scheduled price increases in frontier AI never took effect.
Three vendors could reprice about 21 times by 2031.
Google has averaged one flagship launch roughly every eight months. If OpenAI and Anthropic keep a similar pace, your cost of goods could change about 21 times by 2031. Teams that can switch models with one config change can buy from the cheapest capable model each time.
2031 and the $4/$20 Anchor
Proactive calls Argon Google's first flagship model since February. February to the end of September is about eight months. Five years is 60 months, so at that cadence Google has about seven more flagship launches before 2031. Each one can arrive with its own introductory window and its own shadow price.
My assumption is that OpenAI and Anthropic ship at a similar pace. Seven launches times three vendors is about 21 42 repricing events by 2031. Each is a moment when your cost of goods changes without anyone asking you.
Prices move in both directions. OpenAI priced GPT-6.1 Sol at one-fifth of GPT-6 Astra's standard rates, an 80% 43 drop for what it calls "nearly the same level of intelligence." Google has scheduled the opposite move for Argon: up 100% 25. Business Model Analyst adds that the last two scheduled price increases in frontier AI never happened.
That uncertainty is the argument for building at the shadow. Price your product to survive $4 11/$20 and every outcome works. If Google never raises the rate, the gap becomes margin. If it does, nothing breaks.
Build only for $2 11/$10 and the bet flips. You gain nothing extra if prices hold. You lose your whole margin if they don't, on a date you do not control.
Portability compounds across those 21 42 events. A team that can move traffic with one config change gets to buy from the cheapest capable model every time a sticker changes. A team hard-wired to one endpoint eats each new price. My read is that $4 03/$20 is the honest planning anchor for frontier work through 2027, since Google has posted it and Anthropic already charges it for Opus 5.5.
Rerun Your Unit Economics at $4/$20
First, log four numbers on every request: fresh input tokens, cached input tokens, output tokens and a pass or fail on the task. Cached input means context the provider already holds from an earlier call. It bills at $0.10 09 per million today. If your provider dashboard only shows totals, write the counts into your own request logs.
Then price last month's logs twice in one spreadsheet. Column A uses today's sticker. Column B doubles every rate, which assumes Google keeps the 95% 28 cache discount after the introductory period. Divide each column by the tasks that passed, and you get cost per successful task, the one price your offer actually depends on.
Next, draw a margin line. If column B pushes your gross margin below what your offer needs, fix the output side first. Cap max output tokens, ask for diffs on code edits, and put your long system prompt where the provider can cache it.
After that, move the model name into a config file and pick 20 real tasks from your logs. Run them through two providers. Most readers cannot call Argon yet, so pair Sol with a model you can reach that already charges $4 03/$20, such as Claude Opus 5.5. That run gives you a live preview of Argon's shadow price.
Some of those 20 tasks will break. Calling two models is easy; getting the same result from both is the expensive part. Tool calls will fail on one side and structured output will drift on the other. Each failure is a row in your sheet, and those rows tell you what switching would really cost.
Rerun the sheet on the first of every month, and again the day any vendor posts a new rate. Google will post one for Argon eventually. When it does, you will already know what column B says.
Rerun your unit economics at $4/$20 before Google sets the date.
- Log four numbers per request. Record fresh input tokens, cached input tokens, output tokens and a pass or fail on the task, in your own logs if your dashboard only shows totals.
- Price last month twice. Put today's sticker in column A and double every rate in column B. Divide each column by the tasks that passed to get your cost per successful task.
- Move the model name into config. Run 20 real tasks through Sol and a model that already charges $4 03/$20, such as Claude Opus 5.5. Log each tool call or structured output that breaks as its own row.
Build for $4 33/$20 and let $2/$10 become margin.
The matching price tag shows two labs reading the market the same way, not a floor that will hold. Google has already posted the higher rate, and Anthropic already charges it for Opus 5.5. Price your product to survive the shadow rate, cut output tokens first, and keep switching providers to a one-line change. If prices never rise, the gap becomes profit. If they do, nothing breaks.