The promotional price for Claude Sonnet 5 ends today, 1 September 2026. The price goes from about $2 to $3 per million input tokens, a 50% increase. But the real increase in your bill will be higher, because a tokenizer change makes the model consume up to 35% more tokens for identical text.
We flagged this deadline on 1 August, a month in advance. It is here. Here is what you should do today rather than discovering it on your end-of-month bill.
Why the increase goes beyond what is announced
The point many miss comes down to a distinction we explain regularly. The price is expressed per million tokens, not per million characters. But a token is not a fixed unit: it is a fragment of text whose segmentation depends on the tokenizer used by the model.
We detailed this in our article on tokens. If the new tokenizer splits the same text into more pieces, you pay more for identical work, without the displayed price changing.
The calculation therefore becomes: roughly a 50% price increase, multiplied by a consumption increase of up to 35%. Depending on your texts, the real increase can approach a doubling.
This effect is particularly pronounced for French and languages other than English, which are generally tokenised less efficiently.
Measure on your own texts. Don't rely on an average: count the tokens consumed by a representative sample of your real requests, before and after.
Check your caps. If you have configured monthly spending limits, they could be hit sooner than expected.
Look at the cache. If part of your requests reuse identical context, caching remains the most effective lever, and the cache price differs from that of standard input.
What this says about the market
This increase comes in a contradictory context we have been documenting for weeks.
On one hand, prices are falling: OpenAI's top tier has dropped by more than 20%, and the free offering has become unlimited. On the other, they are rising: DeepSeek has increased by 93%, and here is a promotional end at 50%.
So there is no single trend, and that is the information. A price reflects a commercial strategy — acquisition, retention, monetisation — as much as a production cost.
The general lesson
We have repeated it for months and this deadline illustrates it perfectly: don't build your economics on a promotional price.
Three practices genuinely protect you.
Isolate the model call in your code, so you can switch providers without rewriting the application.
Keep a reference test set, around twenty cases with expected outputs, which lets you measure whether an alternative performs just as well on your usage.
Measure the cost per completed task, not the displayed price per million tokens. It is the only metric that matters, and it incorporates both the price and the model's verbosity.
What to remember
This increase is not a scandal: a promotion ending on the announced date is normal practice, and Anthropic gave plenty of notice.
What deserves attention is the invisible part. A displayed price does not tell you what you will pay, because the billing unit itself can change. That is a lesson that will apply to all future migrations, and there will be more.
If you had done no checks, today is the right day. The wrong day is 1 October, when you discover the bill.