The long-context cliff, and why one extra token can double your bill
A model with a million-token window sounds like a single number: how much it can hold. For several current models, capacity and price are actually two separate lines, and the second one matters more than the first.
What a tier actually does#
A handful of models — Gemini 3.1 Pro and the GPT-5.6 family among them — publish a threshold on input tokens. Below it, the model bills its normal rate. The instant a request's input exceeds it, the entire request — every token of it, not just the tokens past the line — is billed at a multiplied rate. Output is multiplied too, by its own factor.
That "entire request" detail is the part worth sitting with. This is not a marginal rate that only applies to the excess, the way an income tax bracket works. It is a step function: one side of the line, one price; the other side, a different and higher price, applied retroactively to the whole thing.
The numbers, worked#
Gemini 3.1 Pro's threshold is 200,000 input tokens. Below it, input runs $2 per million. Cross it, and input jumps to $4 per million — a 2x multiplier — while output goes from $12 to $18 per million, a 1.5x multiplier.
Take a request at exactly 200,000 tokens and one at 200,001:
- 200,000 tokens: still under the threshold. Cost: 200,000 ÷ 1,000,000 × $2 = $0.40.
- 200,001 tokens: one token over. Cost: 200,001 ÷ 1,000,000 × $4 = $0.80.
One additional token roughly doubled the bill for the request that carried it. The same shape holds on GPT-5.6 Terra, whose threshold sits further out at 272,000 tokens but multiplies by the same 2x input / 1.5x output: a request at 272,000 tokens costs $0.544; one token past it costs $1.088.
Not every large window has a cliff#
This is a property of specific models, not of large context windows in general. GPT-4.1 accepts 1,047,576 tokens with no tier at all — its $2-per-million input rate holds from the first token to the last. Gemini 3.6 Flash is the same story at its 1,000,000-token ceiling: no threshold, no multiplier, flat $1.50 per million regardless of size.
So two models can advertise comparable windows and price a large request completely differently. A 300,000-token prompt costs the same per token as a 300-token one on GPT-4.1. The identical prompt on Gemini 3.1 Pro is already 50% past its threshold and paying double.
Why this is easy to miss#
A per-million headline rate is what gets compared model to model, and it is true — right up until a request crosses a line that headline rate says nothing about. The threshold and the multiplier are separate facts from the base price, published separately, and easy to leave out of a mental model built from "$X per million tokens."
It also compounds with anything else moving the price. A promotional rate, where one is in effect, is what gets multiplied — the tier does not somehow cancel the discount, it stacks on top of whatever rate was already in force.
What to do with this#
Before treating a long prompt as "a bit more of the same," check whether the model you're pricing has a longContextThreshold at all — this site's model pages list one wherever it applies, next to the base rate. If it does, the number that matters is not "am I sending a lot of tokens" but "which side of that specific line am I on," because the cost step happens at one exact token count, not gradually.
And if a workload regularly sits close to a threshold, that proximity is worth knowing on its own: a prompt that grows by one more retrieved document, one more turn of history, or one more page of a pasted file can move the entire request's bill by a fixed multiple, for a reason that has nothing to do with how much bigger it got.