Why comparing model prices side by side is misleading — three hidden multipliers

The dollars-per-million-tokens number tells you almost nothing on its own. Tokenizers, long-context cliffs and expiry-dated discounts change the answer completely.

Bu yazının Türkçesi: Türkçe sürüm.

Today I visited five different "AI pricing guide" sites and got five different prices for GPT-5.6 Sol. One said $5, one said $4, one mentioned a 20% discount but got the date wrong. I gave up and went to the three providers' own documentation. Every number in this post comes from there, as of 25 July 2026.

But the wrong numbers are not the real problem. The real problem is this: "X dollars per million tokens" is not, on its own, a comparable quantity. There are three hidden multipliers and all three live outside the pricing table.

First, the list prices

The starting point. All figures are dollars per million tokens, input / output:

Sources: Anthropic pricing, OpenAI API pricing, Gemini API pricing.

It is very tempting to look at that and conclude "Sonnet 5, Terra and Gemini 3.1 Pro all cost the same." That would be wrong.

Multiplier 1: tokenizers count the same text differently

At the bottom of Anthropic's pricing page there is a note nobody reads. In their own words: Claude 4.7 and later use a new tokenizer, and it produces approximately 30% more tokens for the same text.

Think about what that means. Opus 4.6 costs $5. Opus 5 costs $5. Two identical rows in the price list. But hand the same PDF to both and Opus 5 will have read about 30% more tokens. So the real price of the same job is not $5, it is an effective $6.50.

Anthropic is not hiding this — it is right there in the docs, with a rationale: the new tokenizer contributes to the model's performance. That is probably true, and it may even need less thinking per token. But for someone comparing prices the conclusion is unchanged: the list price held still while the cost of a unit of work went up.

This also muddies cross-vendor comparison. Claude and GPT already tokenize differently; one vendor's "token" is not the same amount of text as another's. Nobody publishes a side-by-side count of how many tokens two providers see in the same document. I can't publish one either, because Claude's tokenizer isn't public. The only honest thing I can say is: the numbers on those pricing pages are in different units, and nobody corrects for it.

Multiplier 2: the long-context cliff

All three vendors advertise context windows around one million tokens. What happens once you actually go in there is completely different:

Note "the entire request". On GPT-5.6, going one token past the threshold reprices the first 272k as well. Not a tier — a cliff.

Let's run the same job in each place. 300k tokens in, 5k tokens out — roughly what feeding a large codebase in one shot looks like:

Now the same models below the threshold, at 50k in and 5k out:

The ordering changed. On the small request Sonnet 5, Terra and Gemini 3.1 Pro sit in effectively the same place; on the big one Sonnet 5 drops to half of the other two. Opus 5 is the most expensive thing on the list for small requests and lands below Sol for large ones.

In other words: the answer to "which is cheaper" swaps around depending on how much context you send. There is no single ranking.

Multiplier 3: prices have expiry dates

Some of the numbers you see today are not permanent. All three vendors state the dates in their docs, and nobody reading the table scrolls to the footnote:

Put those three together and you get an interesting picture: in the same quarter, Anthropic cancelled a planned increase, Google scheduled one, and OpenAI declared its discount temporary. Three different pricing strategies.

The practical consequence: if you are budgeting four months out, use the date in the footnote, not the price on the page. A product built on Gemini 3.7 Flash will see its model cost double on 1 January, and that is not a surprise — it is announced.

So what should you do

My own conclusion came down to three steps:

1. Measure the job, not the price. Instead of "what does a million tokens cost on this model", ask "what does my typical request cost on this model". If you don't know your typical request's input-to-output ratio you cannot compare anything — output costs five times input, and that ratio varies between models.
2. Decide which side of the cliff you live on. If you routinely work above 272k tokens, Claude's flat structure is a serious advantage. If you are always under 50k, that advantage is worth exactly nothing and a Flash-class model will take you far cheaper.
3. Account for caching. All three charge roughly a tenth of base input for cache reads. If you run a fixed system prompt your real bill sits well below list price — but only if you set the caching up properly.

And finally: do not get these numbers from a third-party comparison site. I checked five of them today and every one was wrong somewhere. All three vendors publish their prices publicly and reading them takes ten minutes.

← All posts