The spread
Models of comparable capability differ in price by more than an order of magnitude. Both prices are real, and neither is a quality signal.
Part of the Cheap and fast track on lAItest.
Two models that pass the same tests can differ in price by more than an order of magnitude. Nobody is confused. Both prices are real.
Price is set by strategy, not by cost
A token has no natural price. A lab with the most capable model prices to defend that position. A lab buying share prices to take it. A lab that has already published its weights prices near what serving costs, because anyone can serve the same model. The spread across the top of the market is far wider than the spread in what those models can do.
The same job, priced across the market
| What | Value | Provenance |
|---|---|---|
| DeepSeek V4-Pro input | $1.32 per 1M | source, verified . |
| DeepSeek V4-Pro output | $3.96 per 1M | source, verified . |
| Claude Fable 5 input | $10 per 1M | source, verified . |
| Claude Fable 5 output | $50 per 1M | source, verified . |
| Cheapest frontier input | $0.75 per 1M | source, verified . |
| Median frontier input | $2 per 1M | source, verified . Both rows above are frontier models, tracked in the same index. |
Try it
Sort every tracked model by price, then by context window. The order changes more than you expect. This step is an interactive widget; open the lesson to use it.
A common misconception
Commonly believed: The expensive model is the good one, so price is a rough proxy for capability.
Actually: Rankings are per task and they disagree. On the July 2026 releases, one lab top-scored an agent benchmark while losing to a rival on a software-engineering benchmark, and the pricier of the two was not the winner on both. A benchmark result also only means something with the harness named, because different scaffolds are not comparable.
In one sentence
There is no market price for a token. There is a price per vendor per strategy, and knowing the spread is part of your job.