The price rise already on the calendar

Gemini 3.7 Flash and Gemini 3.6 Flash are currently priced at $0.75 per million input tokens and $3.75 per million output tokens. That is introductory pricing, and it is published with an end date.

On 1 January 2027 those rates become $1.50 and $7.50. The same models, the same calls, twice the price.

This is not a forecast. It is on Google's own pricing page with a date attached, and most cost models built this year do not contain it.

What it does to a real budget

Take a document pipeline running 40,000 pages a month, at roughly 1,800 input tokens and 400 output tokens per page:

input:   40,000 × 1,800 =  72,000,000 tokens
output:  40,000 ×   400 =  16,000,000 tokens

At today's rate:

72 × $0.75  = $54.00
16 × $3.75  = $60.00
              -------
              $114.00 / month

From 1 January 2027:

72 × $1.50  = $108.00
16 × $7.50  = $120.00
              --------
              $228.00 / month

At this scale the change is $114 a month and nobody notices. Multiply the volume by fifty — a mid-sized ingestion product — and it is $5,700 a month appearing in January against a budget written in August, with no corresponding change in usage. That is the version that turns into an awkward meeting.

Why it matters more than the number

The dollar amount is usually survivable. What breaks is anything with a fixed price on the other side of it.

If you sell document processing at a per-page rate, or a per-seat SaaS plan with AI features included, your cost of goods doubles on a date you did not choose, and your revenue does not move. A contract signed in October at a margin that assumed $0.75 input is a contract that quietly loses a chunk of its margin in January.

Three places to check before signing anything that runs past the new year:

1. Per-unit pricing you have already published. What is the margin at 2× inference cost? If it goes negative, that is a repricing decision with a deadline on it.
2. Annual contracts spanning the date. Twelve months at one price against a cost base that steps up in month four.
3. Free tiers and trials. These are pure cost. A generous trial that was cheap becomes twice as expensive on the same day.

The general lesson, which outlasts this particular date

Introductory pricing is normal and it is not a trap — it is how providers seed adoption of a new model. The mistake is treating a promotional rate as the permanent one because that is what the API returned when you built the spreadsheet.

Two habits fix it permanently:

Record the rate with a date. In the usage ledger, price each row when you write it and keep the rate table versioned with a "checked on" date. Then a rate change shows up as a step in a chart rather than silently rewriting your history.

Model the standard rate, not the promotional one. When quoting a project that will still be running next year, build the estimate at the post-promotion price and treat the discount as temporary upside. It is the same discipline as not budgeting on a first-year cloud credit — the credit is real, and it is not the run rate.

What to do about it

In rough order of effort:

• Re-run your cost model at 2× on Flash. Ten minutes. Do this first, because it tells you whether anything else on this list is worth doing.
• Check whether Flash is even the right tier. The gap between Flash and the cheaper or more capable tiers changes when Flash doubles, and a routing decision made under introductory pricing may no longer hold. So might the comparison against other providers — Claude Haiku 4.5 at $1/$5 is currently more expensive than Flash and becomes the cheaper of the two in January, which is the kind of reversal that only shows up if you re-run the comparison.
• Look at batch and caching before you look at models. Google, Anthropic and OpenAI all discount asynchronous batch processing by around half. For anything without a latency requirement that is a larger, more certain saving than a model change, and it does not require re-validating quality.
• Put a calendar reminder on 15 December 2026. Not for the price rise itself — for checking whether the date moved. Anthropic cancelled a scheduled increase on Claude Sonnet 5 that had been announced for 1 September 2026; the introductory $2/$10 became the standard price instead. Announced dates are not guaranteed in either direction.

That last point is the honest caveat on this entire article. Providers change their minds, usually downward, as inference costs fall. Plan for the published date, and check it before you act on it.

More in when the model bill triples, which feature tripled the bill, and what an AI feature costs to build. Pricing quoted from Google's and Anthropic's published rates, checked 27 August 2026.