Gemini 3.7 Flash Pricing Doubles on January 1: The Promotional Rate Budget Trap

The Gemini 3.7 Flash price your budget rests on, $0.75 per million input tokens and $3.75 per million output tokens, expires on December 31, 2026. From January 1, 2027 Google's pricing page lists $1.50 and $7.50, and every batch, flex, priority and caching rate steps up by the same 2x. Gemini 3.7 Flash pricing is not a price. It is a promotion with a date on it, and the model itself was deprecated on October 8.

What Google's pricing page actually says

The short version first. Gemini 3.7 Flash pricing was $0.75 per million input tokens and $3.75 per million output tokens, described by Google as an introductory price through December 31, 2026. On January 1, 2027 the listed rate becomes $1.50 and $7.50. Requests to 3.7 Flash are now served by Gemini 3.8 Flash, which carries the identical dated schedule.

Standard rates and caching

Input goes from $0.75 to $1.50 per million tokens. Output, which on this page includes thinking tokens, goes from $3.75 to $7.50.

Caching does not escape. A cache read goes from $0.075 to $0.15, and cache storage goes from $0.50 to $1.00 per million tokens per hour. The discount you engineered with long cached prefixes now applies to a number that has doubled.

Batch, Flex and Priority

Batch and Flex sit at half of Standard, so they go from $0.375 and $1.875 to $0.75 and $3.75. Priority sits at 1.8x Standard and goes from $1.35 and $6.75 to $2.70 and $13.50. Only the base moves.

One detail the page leaves out and Google's launch post supplies. The August 13 announcement describes the $0.75 rate as "an introductory price of half the original 3.6 Flash cost per million tokens". The 2027 rate is not a new price. It is the price Gemini 3.6 Flash launched at in July, restored.

Gemini 3.7 Flash lasted 56 days as a billable model

Search for Gemini 3.7 Flash pricing today and you will not find it on Google's pricing page. The Gemini API changelog records general availability on August 13, 2026, "available at an introductory price through December 31, 2026", and deprecation on October 8, 2026, with all requests automatically routed to gemini-3.8-flash.

That is 56 days. The deprecations page lists no shutdown date, because none is needed. The model string still resolves. It just bills as, and behaves as, a different model.

The routing note is the part that should worry you

On September 2, when 3.8 Flash shipped, Google's announcement said developers could "continue to rely on Gemini 3.7 Flash, which remains fully supported for efficiency-first workloads". The same post warned that 3.8 Flash "might use more tokens to maximize performance, especially at higher effort levels". Thirty-six days later the efficiency-first option was routed into the model that uses more tokens.

I do not think this was bad faith. Three Flash releases in six weeks, by Google's own count, is a cadence no deprecation policy was written for. The budget effect is the same whatever the intent: a token-count drift you did not choose, under a 2x rate step you did not set.

What this means for the number in your forecast

A model ID is a routing label, not a contract term. Nothing on the pricing page, the changelog or the deprecations page commits Google to a given model at a given price beyond the dated rows. The lock-in mechanics are covered in the piece on model switching costs. The budgeting consequence is narrower. You cannot hedge a dated promotional rate by staying on the old model, because the old model can be removed faster than the promotion expires.

Every dated promotional API rate we could verify

This is the table the post exists for. The inclusion rule is strict: the provider itself has to have published the date, on a page we opened. Aggregator lists that restate a date did not count, and one widely repeated promotion was dropped for that reason.

The GPT-5.6 Sol prior rate is arithmetic, not a quoted figure: AWS describes the cut as 20% on input and 33.3% on output, which only resolves to $5 and $30.

What did not make the table. Z.ai's pricing page shows GLM-5.3-Flash at $0.15 and $0.50 with no promotional wording, although an aggregator reported a dated 50% launch discount to September 9, so it is out. The same page carries "Limited-time Free" against cached input storage with no end date at all, which is the worst case for a budget. DeepSeek is out for the opposite reason: its pricing page halves rates outside weekday peak hours as a permanent schedule, not a promotion. The xAI and Mistral pages we opened carry no dated rate of any kind.

Three kinds of expiry wording, and they are not the same risk

The register looks like one phenomenon. It is three, and each needs a different line in the forecast.

A hard date with a published successor rate

Google's wording is the most useful kind. You know the day and the new figure, so you can book it. Whether Google changes its mind before December 31 belongs in the upside case, not the base case.

A floor date with no successor rate

OpenAI's phrase for GPT-5.6 Sol is "available at least through November 21, 2026", and the launch announcement said "for the next 3 months". That is a floor on the discount, not a date on which anything is scheduled to happen.

The temptation is to read "at least" as a soft promise of extension. I would read it the other way. Three days ago OpenAI announced GPT-6 Sol at $2 and $10 per million tokens, half the promotional rate on its predecessor. The promotion now looks like a bridge for customers who have not migrated, and bridges get removed once traffic has crossed. The defensible assumption after November 21 is the prior $5 and $30, with a migration project costed alongside it.

An introductory rate that became permanent

Anthropic is the counter-example, and a real one. Claude Sonnet 5 launched on June 30, 2026 at $2 and $10 per million tokens as introductory pricing through August 31, with $3 and $15 scheduled from September 1. On August 10 Anthropic edited the launch post and its pricing documentation to say the increase "will not occur".

That is good news for anyone who budgeted at $3 and $15. It is irrelevant to anyone who budgeted at $2 and $10, because they got lucky rather than right. Carry an introductory rate at the scheduled list figure with the discount shown as a dated credit, and a reversal shows up as savings. The wider shape of Anthropic's plans is in the breakdown of Claude pricing plans and API rates.

The step-up on a real workload

Percentages hide the money. Here is a fixed workload run through the published rates: 1 billion input tokens and 200 million output tokens a month. Every figure is arithmetic, not a measurement.

Gemini Flash, tier by tier

On Standard, 1 billion input tokens at $0.75 is $750 and 200 million output tokens at $3.75 is $750, so the bill is $1,500 a month. From January 1 the same traffic bills at $3,000. Batch or Flex goes from $750 to $1,500. Priority goes from $2,700 to $5,400.

The step is $1,500 a month on Standard, or $18,000 a year on a workload most teams would call cheap. On Priority it is $32,400 a year.

GPT-5.6 Sol, if the floor becomes a ceiling

The same workload on GPT-5.6 Sol at the promotional rate is $4,000 of input and $4,000 of output, or $8,000 a month. If the rate reverts to $5 and $30 after November 21, the bill becomes $11,000, a rise of 37.5%. Migrate to GPT-6 Sol at its listed $2 and $10 and the bill is $4,000, before any evaluation and regression cost.

That last clause is the whole decision. The cheapest line on the pricing page is a different model, and a different model is an engineering project with a quality risk attached, the subject of the analysis of the routing quality gap. Price it as a project, not a config change.

Why vendors price this way, and why it works on you

A dated introductory rate does three jobs for the vendor at once. It anchors every third-party comparison table at the low figure for months. It pulls workloads onto a new model before the competitor's next release. And it sets a list price that was never charged, so the eventual step reads as the end of a discount rather than a rise.

The anchoring is the part that costs you. Every comparison we found listed Gemini Flash at $0.75 and $3.75, and most did not mention the date. A buyer who signs an annual plan in November has budgeted at a rate that holds for six weeks of a 52-week year. We have written separately about why token prices fall while AI bills rise.

A second pattern is worth stating plainly. Successor models arrive at the predecessor's promotional price: GPT-6 Sol lists at $2 and $10, half the GPT-5.6 Sol promotion, and Claude Sonnet 5.5 shipped at the $2 and $10 Sonnet 5 was introduced at. My reading is that the promotional figure is becoming the real price of the next model. The standard proposed in the rate limits and pricing transparency piece would require a published successor rate for every dated promotion. Google meets it. OpenAI does not.

Where this argument is weakest

The post says budget at the post-promotion rate. There are four honest reasons that might be too conservative.

First, Google may blink. Anthropic cancelled a scheduled 50% rise with three weeks to go, and Google's competitive context in December is harder, with GPT-6 Sol and Sonnet 5.5 both listed at $2 and $10. A doubling into that field is commercially awkward, and awkward prices get revised. I still would not budget on it.

Second, materiality. On the worked example the Gemini step is $1,500 a month, which for many teams is below the threshold of a reforecast. The argument holds for pipelines at 10 to 100 times that volume and for anyone who priced a customer contract on Flash margins. It is weak advice for a prototype.

Third, tokens per task are not fixed. Anthropic's Sonnet 5.5 announcement claims up to 30% lower cost per task from token efficiency. That is a vendor claim we have not measured, but if it holds, a 2x rate on 30% fewer tokens is a 1.4x bill.

Fourth, this register is a one-day hand census, not a dataset. We opened the official pages of nine providers and found six dated promotions. There are certainly more on regional platforms and inside enterprise agreements we cannot see.

How to budget for a dated rate

Six rules, in the order I would apply them. None needs a new tool.

Keep a promotion register with an owner. One row per model you run: provider, promotional rate, list rate, the provider's exact date wording, and which of the four structures above it is. Review it monthly.

Forecast at list, and show the promotion as a dated credit. The forecast line reads $1.50 and $7.50 from day one, with a negative line called "Gemini introductory credit" that ends on December 31. When a vendor extends, you record a saving. Tooling that automates this is compared in the review of AI spend forecasting tools.

Use fail-closed identifiers where they exist. Vercel's AI Gateway offered Ling 3.0 Flash Fin under two IDs: one that began billing when the free period ended, and a second that returned an error instead. More providers should copy it. Where it does not exist, set a hard spend cap and a billing alert dated to the expiry.

Do not plan to ride out a promotion on the old model. Gemini 3.7 Flash was called fully supported on September 2 and routed away on October 8. If your cost case depends on a model ID surviving, write that into the contract, which is what the CFO's AI contract clause list is for.

Re-measure tokens per task after any forced routing. Google says 3.8 Flash may use more tokens than 3.7 Flash at higher effort levels. A rate step on top of a token drift is two multipliers, and only your own logs show the second one.

Treat "at least through" as the date the discount ends. Put November 21 in the register as the expiry for GPT-5.6 Sol, with $5 and $30 as the rate after. If OpenAI extends, move the date. If it publishes a successor price, replace the assumption.

Frequently asked questions

Where to start this week

One query and one spreadsheet column.

The query: pull token usage by model ID for September, then multiply the Gemini Flash rows by 2 and the GPT-5.6 Sol rows by 1.375. That is your January invoice if nothing else changes. If the number is below what anyone would notice, stop here and set a reminder for December 1.

The column: add "rate after" and "provider's date wording" to whatever sheet tracks your AI vendors, filled from the provider's own page rather than a comparison site. A blank in that column is the model that needs a conversation with the vendor, found before the invoice finds you.

Related reading