The industry story is that tokens keep getting cheaper, and it is half true: each generation of small models arrives cheaper than the last. But 2026 has consolidated four ways for your bill to go up without you changing a single line of code. None of them is a trick: they are all published on the providers’ own pricing pages. The problem is that none of them shows up in the comparison you looked at the day you picked a model.

1. Promotions have an end date

OpenAI currently sells GPT-5.6 Sol at $4 input and $20 output per million tokens. Its pricing page says, literally, that this promotional price is available “at least through November 21, 2026”.

“At least” is a floor, not a ceiling: it may be extended, and it may end on that date. If you budgeted the year on the promotional rate, you have a date in the calendar and an unquantified risk, because what the model will cost afterwards has not been published.

2. Long context is billed in a different tier

This is the one that surprises people most, because it does not depend on the model but on how much you send it.

OpenAI publishes two tiers per model, short context and long context. For GPT-5.6 Sol, input goes from $4 to $8 and output from $20 to $30. Google marks it by threshold: on Gemini 3.1 Pro, below 200,000 tokens input is $2 and above it $4, while output rises from $12 to $18.

The important detail is that OpenAI’s page shows both tiers but not the threshold that separates them. You know there is a different price, but not the exact point at which you enter it. And in practice you get there on your own: a RAG pipeline that keeps accumulating documents, a conversation history nobody prunes, a system prompt that grows one patch at a time. Same task, same model, different unit price.

Anthropic has gone the other way: since Claude 4.6, the one-million-token window is included at standard pricing, with no surcharge. Worth knowing if you are comparing.

3. Uplifts for where and how it is processed

These are configuration flags, not model decisions, and each one has a price.

Data residency. OpenAI charges a 10% uplift on regional endpoints, for models released on or after 5 March 2026 that support it. Anthropic applies a 1.1x multiplier to US-only inference.

Fast mode. At OpenAI it doubles the standard rate. At Anthropic, fast mode on Opus 5 goes from $5/$25 to $10/$50: exactly double.

Neither changes what the model does, only where it runs or with what priority. A misplaced flag in one integration doubles the bill for that route, and nobody notices until the month closes.

4. Rises announced with a date

Google publishes its current prices as valid through 31 December 2026 and has already announced that from 1 January 2027 they roughly double across the line. Gemini 3.8 Flash input goes from $0.75 to $1.50. Context cache storage goes from $0.50 to $1.00 per million tokens per hour.

It is the largest of the four rises and also the easiest to plan for, because it has been published almost four months ahead. It is also the one most people will meet in their January invoice without having looked at it first.

And sometimes the rise never happens

Anthropic had announced that Claude Sonnet 5 would go from $2/$10 to $3/$15 on 1 September 2026. It did not happen: the launch rate has stayed on as the standard price.

That is good news and, at the same time, the best argument in this article. The price of tokens is a variable, not a constant: it goes up, it goes down, and sometimes it is announced and withdrawn. Finding out by reading is cheap. Finding out from the invoice is not.

What to do about it

Store the expiry date next to the price. In your cost sheet, every rate should carry where it came from, when you verified it and how long it is valid. A price with no date is a figure that expires in silence.

Measure cost per million tokens, not total spend. If your usage grows 20% and the price rises 15%, total spend goes up and you cannot tell how much is which. Unit cost separates them.

Check the flags before the models. Long context, residency and fast mode are three ten-minute checks that explain rises that otherwise look inexplicable.

The structural discounts are still there. Cache reads cost a tenth of input at both OpenAI and Anthropic, and batch processing is 50% off at all three providers. A large share of the real saving lives here, without changing model or provider.

Switching provider is not switching model. The same model served from somewhere else can carry a different price, a different residency and different uplifts. They are two separate decisions and it is worth not mixing them.

Frequently asked questions

Have token prices gone up in 2026? List prices at the small end have come down, but tiers and uplifts have appeared that raise the effective cost: long context, data residency and fast mode. On top of that, Google has announced a rise of roughly double for 1 January 2027.

What is the long-context tier? A different per-token price once the request goes over a certain size. On Gemini 3.1 Pro the threshold is 200,000 tokens and input goes from $2 to $4. OpenAI publishes both tiers but not the threshold between them.

How do I know if it is affecting me? Look at cost per million tokens month by month, split by model and by route. If it rises without you having changed model, the cause is usually a tier or an uplift, not the provider.

This piece is part of the cluster on AI cost per task. If you want to run your own numbers, there is the token cost calculator.

Prices verified on 7 September 2026 on the official OpenAI, Anthropic and Google pricing pages. They change often: always check the current rate before deciding.

At e-ficient we measure cost per task for companies already running AI in production. The initial audit is free if you start a plan with a three-month commitment and returns a diagnosis within 72 hours.