One model, many sellers: how far token prices spread

How many models have more than one seller

Of the 467 models we hold token prices for, 93 are sold by more than one seller, across 16 providers in all. For each of those models a buyer is choosing between sellers of the same weights, and the per-token rate is the first thing that separates them. A seller of a popular model is therefore priced against its rivals line by line, in public.

This post does not give a spread for any single model. We have no live figure for one yet, and a typed-in number would be wrong by the next read.

What the token table shows a buyer

The token price table puts every seller of a model on one view, with input and output rates per million tokens and the page each was read from. The LLM price index tracks the market across models. A buyer looking at a multi-seller model sees the cheapest and dearest sellers at once.

Where the gap is small, the buyer picks on something else. Where it is large, they ask what the dearer seller is selling. Both questions land on the seller's own page, and the page has to answer them.

A buyer choosing between sellers of the same model may well have tested the model already. The question of whether the model is good enough is settled. What remains is the cost, the limits and the operating terms, which is a narrower, more commercial question than the one a first-time buyer asks.

That buyer switches more easily too. Moving between sellers of the same model is a change of endpoint and key, not a migration, and some route traffic across several sellers at once. Each rate change is read by buyers who could leave in an afternoon.

What a higher rate might be buying

Sellers of the same model can differ in context window served, numeric precision, rate limits, regions, data retention terms, and whether cached or batch tiers exist. Any of these could justify a higher rate. None of them is visible in a per-token figure, and we do not record most of them.

Precision is easy to overlook and changes what is being sold. Weights served in 8-bit or 4-bit take less memory than in 16-bit: about one byte or half a byte per parameter instead of two. A seller serving a lower precision has a cost reason to charge less. If you serve full precision and charge more for it, say so on the row.

Retention is the other quiet one. Whether a seller keeps prompts and outputs, for how long, and whether it trains on them can matter more to a regulated buyer than the rate, and it belongs on the pricing page or one click from it.

The GPU market has the same shape

GPU rentals show the pattern with figures we can fill. On the H100 SXM 80GB the lowest listed hourly rate is $1.68 and the highest is $6.16, the count of sellers is 13, and the median spread across cards is 3.7×. Tokens are produced on GPUs, so there is no obvious reason for token prices of one model to sit closer together than the hardware under them.

That comparison is a reason to expect a spread, not a measure of one. The GPU price index holds the hardware side, and the token side waits for a per-model figure.

Why the cheapest row draws questions

One seller well below the rest raises the question of what was traded for the rate: a shorter window, a lower precision, tight limits or a promotional rate that ends. Sometimes one of those is the answer. When none is, the cheap seller can lose buyers it had already won on price, because it did not answer a question the gap raised.

The dearest row draws the opposite question: what is this for? Paying more makes sense for a seller in the buyer's own region, with retention terms their lawyers accept, or with limits that fit their launch. It does not make sense for a reason the buyer has to guess.

Sellers in the middle raise neither question. For them the facts on the page are what can move them up a shortlist, since the rate alone does not separate them from their neighbours.

What a seller of a popular model can do

Name the model exactly, with maker and version, so the buyer sees your row next to your rivals and not under a separate name. Publish both rates per million tokens. State the served context window and the precision. Put limits by tier on the same page.

If you are the cheapest, those facts stop a buyer assuming the low rate hides a catch. If you are not, they are the main argument you have. The inference API category is where buyers browse the sellers, and a claimed listing lets you answer its questions with a source.

Watch the rows next to yours. When a rival changes its rate, buyers see it on the table the next time it is read, and so can you. Prices older than 21 days are hidden rather than shown stale, so a seller who stops publishing drops out of the comparison entirely.

Sellers of the same model can only see this view from the outside. A claimed listing adds the inside: thirty-day totals of how often a product is viewed, compared and clicked through, which is a first sign of whether the rate is being noticed.