Gemini Flash doubles on 1 January, and the 200K cliff is worse

The $0.75 per million input tokens you are paying for Gemini 3.7 Flash is a promotional rate. It ends on 31 December 2026, after which input goes to $1.50 and output goes from $3.75 to $7.50. Gemini 3.6 Flash reverts on the same date. Google states this plainly on its pricing page; what it does not tell you is what that does to your bill, or that a second and much sharper increase is already live for anyone running long prompts.

We ran the change through three workloads. Every Flash bill doubles exactly:

Computed with our LLM cost calculator against rates verified on 24 August 2026. The exact 2.00x across all three is the point: the promo cuts input, output and cache reads by the same half, so no workload shape escapes it. There is no clever cache ratio that softens this.

Does the revert change which model you should use?

For most workloads, no. For cost-sensitive ones, yes, and by a lot.

At the promotional rate, 3.7 Flash sits mid-table on price. After 1 January it ranks 13th of 31 models we track on the coding-agent workload. Five models are cheaper by more than 5x:

That is not an argument for switching. Nano and Lite tiers are weaker models and a coding agent is exactly where capability earns its price. It is an argument for re-running your own numbers, because a decision made in August against a promotional rate was made against a price that will not exist in January.

The bigger problem: the 200K cliff on Gemini 3.1 Pro

While checking the Flash rates we found something worse, and it is live now rather than in January.

Gemini 3.1 Pro Preview charges $2 per million input tokens for prompts up to 200K, and $4 above it. Output goes from $12 to $18. Our table previously carried a note saying Google had not published the higher figure. It has now, so we filled it in, and the result is not a smooth step.

190,000 tokens in  ->    $892/mo
200,000 tokens in  ->  $4,540/mo
210,000 tokens in  ->  $4,740/mo

Same workload, 5,000 requests a month, 6K output, 80% cache hit rate. A prompt 5% larger costs five times more.

Doubling the rate explains a 2x jump. The rest is the prompt cache. Google publishes a cached-input rate of $0.20 per million for Gemini 3.1 Pro below the threshold, and publishes no cache-read rate at all for the tier above it. Our calculator bills unpublished cache reads at the full input rate, on the principle that a rate you cannot cite is not a discount you can bank. So crossing 200K does two things at once: it doubles the headline rate, and on our modelling it voids the 80% cache discount that was doing most of the work.

Be clear about which half of that is fact. The $4 and $18 tier rates are published and verified. The cache behaviour above 200K is our assumption in the absence of a published figure, and it is the conservative one. If Google is quietly applying the same $0.20 cache rate above the threshold, the real number is closer to $1,800 than $4,540. We have not found a source either way, and we would rather show you a ceiling with the reasoning attached than a comfortable number we cannot defend.

Either way, the actionable part does not depend on the assumption: if your prompts sit anywhere near 200K, measure them. A retrieval step that grows with your corpus, or an agent that accumulates conversation history, will cross that line on its own one day and the bill will move by multiples rather than percentages. Our token counter will tell you where a given prompt actually lands.

What we changed in our own data

This article exists because the check changed our table, not the other way round. Two corrections, both in src/lib/llm-models.ts:

1. Gemini 3.7 and 3.6 Flash carried $0.75 and $3.75 as standard rates. They are promotional. The standard rates are now recorded as $1.50 and $7.50 with a promo block expiring 31 December 2026, so the calculator quotes the right number depending on the date you are planning for.
2. Gemini 3.1 Pro Preview now carries its real long-context tier instead of a note saying the figure was unpublished.

The first one matters beyond this article. Anyone budgeting 2027 spend against a calculator that treats a promotional rate as permanent, ours included until today, was underestimating Flash by half.

What we did not check

Honesty about scope, since this was a Gemini-focused pass:

• Anthropic's pricing docs returned a 404 at the URL we track, and OpenAI's pricing page returned a 403 to our fetch. Neither provider's rates were re-verified in this pass, so if you are comparing across providers, treat the non-Google rows as verified on the previous date rather than today.
• Whether the promo revert applies to batch. Google publishes a flat 50% batch discount, and we assume it applies to the post-revert rate rather than the promotional one, but the pricing page does not spell this out.
• Whether 3.6 Flash survives to January at all. A model on promotional pricing three months before a price rise is a reasonable retirement candidate, and Google has not said.

If you are running any of this in production, the thing worth doing this week is not switching models. It is opening your own numbers, setting the date to January, and seeing whether the answer you got in August still holds. The cost calculator prints the date its rates were verified for exactly this reason, and as the economics of AI infrastructure keeps demonstrating, introductory pricing is a customer acquisition line item rather than a permanent state. The same caution applies to every new model launch: the launch price and the price you will pay in a year are different numbers.