One token can double your bill

Send a prompt of 272,000 tokens to OpenAI's largest model and the input costs $1.09. Add one more token and the input costs $2.18. Not $1.09 plus a fraction of a cent — exactly double, because the surcharge applies to the whole request rather than to the tokens above the line.

OpenAI's model page says so plainly: "Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request."

What the step actually costs

Take a request with 10,000 output tokens, and compare it against Anthropic's largest model, which charges a flat $5 per million in and $25 out.

One token moves the bill by $1.19. At the long-context rate an ordinary input token costs eight ten-thousandths of a cent, so that single token is charged like about 148,000 of its neighbours.

⚠️ Note the shape. This is a step, not a slope. There is no band where the two vendors converge and then cross — the price jumps, and it jumps past Anthropic in one move. Above the line OpenAI is dearer on input and on output, so no amount of adjusting your output length brings it back.

Anthropic went the other way, deliberately

The same page structure at Anthropic contains the opposite commitment. Their pricing documentation states that the full million-token context window is available at standard pricing, and spells out the consequence: "A 900k-token request is billed at the same per-token rate as a 9k-token request."

Both vendors face the same problem: long prompts are expensive to serve. One priced that cost into a threshold; the other absorbed it. That difference in philosophy matters more than the specific numbers, which will change.

The inversion is real and it is not general

The tempting headline is "which flagship is cheaper flips at 272,000 tokens". That is true of exactly one pairing out of three.

So the surcharge changes who you should buy from only at the top of the range. In the middle it reinforces a decision you had already made, and at the bottom it is not enough to matter. If you are going to act on one line of this, it is that the cliff is a top-of-range phenomenon.

The complication that makes all of this approximate

Per-token price comparisons only work if a token is the same thing everywhere. It is not.

Anthropic's own pricing documentation carries this note: "This tokenizer produces approximately 30% more tokens for the same text." It applies to Claude 4.7 and later, which includes Opus 5, and the page is explicit that the exact increase depends on the content. We reported the same change when the Sonnet 5 price rise was withdrawn, and the arithmetic above inherits the problem: a document that is 272,000 tokens to OpenAI is roughly 353,600 to Anthropic.

Redo the input comparison on the same text rather than the same token count:

The finding survives the correction and loses about half its margin. It would only fail if the tokenizer difference reached 1.6×, which is well outside the figure Anthropic publishes. The tokenizer difference, which looks like a rounding detail, is doing half the work here.

⚠️ The note names Claude 4.7 and later, and says Sonnet 4.6 and earlier use the previous tokenizer. Haiku 4.5 is an earlier model, so the correction does not apply to the bottom row of the table above. We are following the documentation's own line rather than assuming it generalises.

Two other cliffs, one token apart

Google and xAI both put a long-context threshold at the same headline figure of 200K, and word it differently. Google's applies to prompts above 200k, so the first surcharged request is 200,001 tokens. xAI's applies when the prompt reaches 200k, so the first surcharged request is 200,000.

The same nominal threshold begins one token apart. This is not worth losing sleep over, but "200K" is a marketing figure and the real threshold sits in the comparison operator.

What to do about it

If your prompts sit anywhere near a threshold, measure them rather than estimate. A retrieval step that pulls "about 250,000 tokens of context" is a step whose output varies with the corpus, and the difference between 271,000 and 273,000 is a doubled bill on every request that crosses.

The cheap defence is a hard cap below the line, enforced before the call rather than discovered on the invoice. Our token counter and tokenization visualizer will tell you what a given text actually comes to, and the LLM cost calculator and price comparison work from published rate cards. None of them models the surcharge. A tool that multiplies tokens by a rate will show a 272,001-token request costing one token more than a 272,000-token one, missing the cliff entirely.

Where this comes from, and what will date it

The surcharge sentence and the per-million prices were read from OpenAI's own model page. The flat-rate commitment, the model prices and the tokenizer note were read from Anthropic's own pricing documentation. Every figure above is arithmetic on those two pages and reproduces in a spreadsheet.

⚠️ This will date, and OpenAI has already published one expiry date. OpenAI's page states that the $4 and $20 rates are promotional and available at least through 21 November 2026. The inversion depends on them: it exists because $4 doubles to $8 and lands above Anthropic's $5. If the promotional rate ends and list pricing is higher, the arithmetic changes and may well disappear. Check the date on this piece against today's before acting on it.

Two limits apply here. First, we have not tested the tokenizer ratio on our own corpus; we are using Anthropic's published 30% figure, which their page says depends on the content. Second, we compared first-party API pricing only, not the partner platforms, which set their own rates and are a different calculation.