The model didn’t change. The invoice will. Who in your company knows the date?
Google’s Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens through 30 November 2026. From 1 December it costs $4 and $20. Same model, same tokens, double the bill. Those are the dates and prices on Google’s own pricing page.
If anyone in your company built something on Argon at the launch price, this post is for you. There are three weeks left to plan.
💸 The Old Price and the New Price
All figures below are USD per million tokens, from the pricing tabs on Google’s Gemini API pricing page. (Google Gemini API pricing)
Standard
– Input: $2.00 through 30 November, then $4.00 from 1 December
– Output (including thinking tokens): $10.00, then $20.00
– Cached input: $0.10, then $0.20
Batch and Flex (both listed at the same price)
– Input: $1.00, then $2.00
– Output: $5.00, then $10.00
– Cached input: $0.05, then $0.10
Priority
– Input: $3.60, then $7.20
– Output: $18.00, then $36.00
– Cached input: $0.18, then $0.36
Look at the pattern. Every tier on the page doubles on the same day. Choosing Batch or Flex gives you a cheaper bill on 1 December, but it does not save you from the rise.
The page also says Argon is not available on the Free tier. So this is a paid-tier story from the start. (There is a separate Contributor row at $1 input and $5 output with no date, described on the page as a private preview offered on an invite-only basis. It is not something you can plan a budget around.)
📌 The Footnote That Was There From Day One
📌 Real-world case: Google’s launch post says Argon “will launch at an introductory price” of $2 and $10, and a footnote says: “After the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply.” The launch post itself gave no end date. The pricing page now does: 30 November. (Google blog)
So nobody was misled. The doubling was written down on launch day. What arrived later was the date.
That is how a lot of price rises work. The rise sits in a footnote, the date lands in a different document, and the team that signed up on launch day has moved on to the next project.
🧮 What It Does to One Workload
Let me use a made-up but typical agent task. This is arithmetic on Google’s published rates, not a measured workload, so swap in your own numbers.
Say one run reads 200,000 input tokens and writes 60,000 output tokens.
- At the introductory Standard price: 200,000 ÷ 1,000,000 × $2 = $0.40 for input, plus 60,000 ÷ 1,000,000 × $10 = $0.60 for output. That is $1.00 per run.
- At the new Standard price: $0.80 plus $1.20 = $2.00 per run.
- At 1,000 runs a month: $1,000 becomes $2,000.
Now look at where the money goes. Output is about 23% of the tokens in this run (60,000 of 260,000) but 60% of the introductory bill. That share stays at 60% after the rise, because both prices double. If your tasks write a lot, cutting output length is the first lever.
The same run at Batch prices is $0.50 now and $1.00 from 1 December. That is half the Standard bill, but Batch is for work that can wait, not for a live chat window.
🧑💼 Why Finance Finds Out Late
Sach batao: who in your company would notice a price change on a model used by one team?
Usually it goes like this. A developer picks a model and writes the name into a config file. The API key sits under one project. The invoice arrives weeks later as a single line. By then the launch price has quietly become “the price” in everyone’s head.
None of that is anyone’s fault. It is how small experiments become cost lines. But a date like 1 December is exactly the kind of thing that should be on a calendar before it happens, not discovered on an invoice.
📖 How to Read a Pricing Page for Dates
The Argon rows are a good example of what to look for on any AI pricing page. It takes about two minutes.
- Find the date column or the footnote. If a price has a “through” date, it is temporary. Write it down next to the price, not in your head.
- Read every tier, not just Standard. On this page, Batch, Flex and Priority all carry their own dates. A saving on one tier can hide a rise on another.
- Separate input from output. The two prices rise together here, but your usage will not. A summarizing tool reads a lot and writes a little. An agent that drafts and re-drafts does the opposite.
- Check the cached price too. If your prompts repeat, caching matters. Its price also doubles on 1 December, from $0.10 to $0.20 on Standard.
- Save a dated copy. A screenshot or a PDF of the page, with the date you read it, settles arguments later about what the price was supposed to be.
This is not about distrusting a vendor. It is about treating a price as something with an expiry date, the way you would treat a quote from a supplier.
🪙 Token Truth
A launch price is a loan, not a rate.
The price per token is only half of the cost. The other half is the date it applies from. A number without a date is a guess about next quarter. (If you want the wider story of how AI vendors price, limit and profit from tokens, start with Episode 1 of The Token Diaries and then Episode 5, Selling by the Syllable, which covers scheduled price rises more broadly.)
✅ Five Things to Do Before 1 December
- List every workload that calls Argon. Who owns the key? Which project? Which team would notice first?
- Re-price each one at the new rate. Use last month’s token counts, input and output separately, and multiply by the $4 and $20 figures.
- Re-run a quality check against the alternatives you already use. Do it on your own tasks. I am not saying any model is better; I am saying you should compare on your own work, now that the price is no longer a reason to skip the comparison.
- Decide what to do about it. Stay, send part of the traffic elsewhere, move patient work to Batch, or cap output length. Each is a legitimate answer.
- Put 1 December in the finance calendar and set a budget alert on the project, so the next surprise is a notification and not an invoice.
Seedhi baat: none of this needs a big meeting. A spreadsheet with five rows is enough.
One caution on step 3. Switching models is not free either: prompts that work well on one model sometimes need rewriting for another, and that engineering time belongs in the comparison. A price rise of 2x does not automatically make a switch worth it. It only means the question is now worth asking.
🔍 What I Could Not Confirm
Two things are worth checking yourself.
First, access. Google’s launch post described a staged rollout, starting with paid API customers and Google AI Ultra subscribers, and the reports I read differ on how quickly accounts got access. Check your own console and the model list before you assume Argon is available to you.
Second, the date can change. It is Google’s page, and Google can edit it. I read it on 7 November 2026. Check it again before you act, and treat anything older, including this post, as a snapshot.
Also be careful with third-party price trackers. I came across at least one that listed different figures from Google’s page. When the two disagree, Google’s page wins.
None of this is financial advice. It is my reading of a pricing page.
If your AI bill doubled next month, would you find out from a calendar or from an invoice?
📚 References
- Google, Gemini API pricing (Gemini 4 Argon rows: Standard, Batch, Flex, Priority; read 7 Nov 2026)
- Google, Gemini 4 Argon launch post (introductory price and footnote)