A price list used to describe the present. Google’s now describes the next ten weeks: seven entries on its Gemini pricing page change on a stated future date, and the new number is exactly twice the old one every time. If you costed an automation this autumn, the rate you used has a shelf life — on one model, five weeks of it.
01 — What happened
Google’s Gemini API pricing page, last updated on 26 October, lists seven models whose published price changes on a dated schedule. The wording is plain: a figure “through November 30, 2026”, then a different figure “starting December 1, 2026”. The later price is double the earlier one in every case.
Gemini 4 Argon goes first. Google unveiled it on 30 September and made it broadly available on 13 October at $2 per million input tokens and $10 per million output. From 1 December those become $4 and $20. Context caching — the standard way to stop paying twice for the same instructions — doubles with them, from $0.10 to $0.20 per million.
The other six move on 1 January 2027: Gemini 3.6 Flash and 3.8 Flash go from $0.75 and $3.75 to $1.50 and $7.50, and two 3.8 Flash speech variants and two Gemini Robotics ER 2 previews double with them. Batch, Flex and Priority tiers carry the same dated changes.
The other large vendors publish something different again. OpenAI lists GPT‑5.6 Sol at $4 and $20, describing that as promotional pricing “available at least through November 21, 2026” — a floor on how long the promotion runs, not an end date, and no successor price. Anthropic’s page is the third variant: a flat list with no promotional rates and no dates on it, under a standing note that price and plans are subject to change at its discretion.
02 — Why it matters
For the last two years the interesting question about AI pricing was the level. This is about the term.
When you cost an automation, you take the current published rate and multiply. That quietly assumes the rate is a fact. On seven Google models it is not — it is the first half of a schedule. A twelve-month business case written in October against Argon’s launch rate is wrong from week six, by a factor of two on its largest variable line.
The honest counterweight: on most workloads this is a rounding error, and nobody is being deceptive — Google published the increase seven weeks ahead and named the figure, which is more than most software suppliers manage. What matters is knowing which of the two you are, and that is not decided by how important the automation is or how many people use it. It is decided by one piece of arithmetic most businesses have never done.
03 — The economics
Take an illustrative business processing 2,000 supplier documents a month through a model that reads each one and returns structured fields.
Assume each document consumes roughly 20,000 input tokens (the document, the instructions, the output schema) and produces 1,500 output tokens. That is 40 million input and 3 million output tokens a month: $80 plus $30 at Argon’s current rate, so about $110 a month, rising to about $220 on 1 December. An increase of $1,320 a year. That is not a decision. It is a line item.
Now change one assumption and nothing else. Instead of one model call per document, the work runs as an agent: it reads, checks a field against a second source, retries a bad extraction, then writes the result. Say twelve calls per document, and because the context accumulates with each pass, an average of 25,000 input and 800 output tokens per call. That is 300,000 input and about 9,600 output tokens per document — roughly 70 cents now, $1.39 from 1 December. At 3,000 documents a month, the line moves from about $2,100 to about $4,180. Call it $25,000 a year, on the same doubling that cost the first business $1,320.
Volume is not what separates them. The multiplier is context length times number of passes. Most teams know their volume and have never measured the other two, which is why hours saved is the wrong way to value an automation.
Then the case that actually hurts. If you charge for that unit of work — per document, per resolution, per qualified lead — you set a price against an input cost with an expiry date on it. At 70 cents of model cost against a $2 price, the gross margin is about 65%. At $1.39, about 30%. Your product did not change, your customers noticed nothing, and the usual mitigation — caching the repeated context — doubles on the same day. This is the mirror image of the shift we covered when software vendors began charging per resolution rather than per seat: anyone reselling metered work carries an input price they do not set.
One number complicates the story usefully. On OpenAI’s own list the newer flagship is the cheaper one: GPT‑6 Sol at $2 and $10 against GPT‑5.6 Sol at $4 and $20. On Anthropic’s, the current Haiku 5.5 is listed at a tenth of the legacy Haiku 4.5. The falling frontier price we wrote about earlier this month is real and still falling. It just does not arrive on your invoice by itself. It arrives if you can move.
04 — What businesses should do
Work out what one unit of work costs you. Divide last month’s API spend by the number of things it processed — documents, enquiries, tickets. One number, ten minutes. Everything below depends on it, and the businesses that cannot produce it are the ones that get surprised.
Check the dates on the models you actually call. Two matter immediately: 1 December for Gemini 4 Argon, 1 January for the Flash and Robotics entries. If your automation names a model in code, someone should know which one and what it costs in February.
If you sell the output, re-run the margin at the post-expiry rate. The only urgent item here. Re-price, absorb it, or change the implementation — but decide now, not in a quarterly review when the fix is a price rise your customers will remember.
Measure passes and context, not just volume. If a unit of work takes twelve model calls, ask what the first eight are doing. Cutting the passes, or what each one carries, reduces the bill more reliably than switching vendor — and makes the automation faster.
Make the model a setting rather than a decision — then do not switch on price alone. The model name should appear in one place. If it appears in six scripts and two automation tools, your response to any price change is a project rather than an edit, the same exposure as when vendors retire whole products. Re-testing prompts against a new model still costs real time, and where the increase is a thousand a year it will cost more than the increase.
05 — Where technology fits
For most businesses reading this, the answer is: nothing to build. This is a costing exercise — one division, two dates, a margin check. Nobody needs a system to do it, and a vendor proposing one in response to a price change is selling you something.
Where there is a genuine technical fix, it is small and dull. If a model name is hard-coded across several scripts, workflow tools and a web application, the fix is one place that decides which model gets called, and everything else asking it. That is an afternoon of integration work, not a platform. The point is not that you intend to switch; it is that switching stops being an event.
If your automation is loop-shaped and you sell its output per unit, the implementation is also a commercial question: fewer passes, shorter context and a cheaper model for the mechanical steps are margin decisions, the same discipline that decides which cases an automation should handle at all.
And the legitimate no: if AI reaches your business only through products other people run — the agent inside your CRM, the summariser in your support tool — you do not have this problem. You have the vendor’s version, which arrives as a packaged price change and is handled at renewal.
06 — RevenueStack’s take
Our view: the development is not the increase. It is that AI pricing has acquired a term structure, and businesses are costing permanent systems against temporary rates.
That exposure is quiet. A subscription increase arrives as an email from a supplier and a renewal conversation. A metered input price that doubles on a published date arrives as nothing at all — no email, no invoice line that reads differently, just a slightly worse month that gets attributed to volume. The businesses that notice are already tracking cost per unit of work. Everyone else feels it as margin drift.
The conclusion is almost the opposite of the obvious one. The answer to prices that move is not to pick the cheapest model, or lock something in, or wait for the market to settle — it will not. It is portability: being able to change the model behind a working process in an afternoon, without renegotiating anything or rewriting much. Falling AI prices are a property of the market. Whether they reach your bill is a property of your architecture.
That is unglamorous to invest in, and it compounds. Every vendor announcement for the next three years — cheaper model, better model, retired model, doubled price — is either an opportunity or an incident, and which one it is was decided long before the announcement.
Sources
- Gemini Developer API pricing — Google, 2026-10-26
- Google makes Gemini 4 Argon broadly available with limited-time low API pricing of $2/$10 per 1M token — VentureBeat, 2026-10-13
- API pricing — OpenAI, 2026-10-27
- Claude pricing — Anthropic, 2026-10-27
