Per-token price is rent, not cost
In August 2026 the cheapest model on the board at $0.06 per million tokens costs $1.409 to answer a thousand exam questions. A model that is not the cheapest per token answers the same thousand for $0.100. Fourteen times cheaper, on list prices anyone can read. Budgets built on per-token rates pick the wrong model, and they miss the trend as well.
Almost every budget I have seen for language-model deployment, including the ones written by people who are careful about everything else, is built on dollars per million tokens. It is the number vendors publish, so it is the number that ends up in the spreadsheet. It is also a rental rate, and what you actually want to know is what a job costs.
Those two come apart, and over the past year they have come apart quickly, because models now decide for themselves how much to think before answering.
The selection error
Take a fixed task: a thousand graduate-level science questions. For every model that clears a quality bar on that benchmark, count the tokens it actually spends answering, and multiply by its own published prices. Then compare the model with the lowest per-token price against the model with the lowest cost for the job.
| Month | Lowest price per token | Its cost per 1,000 questions | Output tokens per question | Actually cheapest for the job | Its cost | Ratio |
|---|---|---|---|---|---|---|
| 2026-03 | Qwen3.5 4B, $0.060 | $0.625 | 4,115 | the same model | $0.625 | 1.00× |
| 2026-05 | Qwen3.5 4B, $0.060 | $0.625 | 4,115 | DeepSeek V4 Flash | $0.151 | 4.13× |
| 2026-06 | Qwen3.5 4B, $0.060 | $1.409 | 9,344 | DeepSeek V4 Flash | $0.145 | 9.75× |
| 2026-08 | Qwen3.5 4B, $0.060 | $1.409 | 9,344 | DeepSeek V4 Flash | $0.100 | 14.12× |
Prices as published by each provider. Token counts are what the models actually emitted on the benchmark.
In March the two answers agreed. Five months later, choosing on rent costs fourteen times the bill of choosing on the job. Nothing about the cheap model's price changed. What changed is that it started spending nine thousand tokens per question instead of four.
The curve that looks flat
The selection error is the visible half. The invisible half is worse, because it distorts the trend rather than one decision.
From March to August 2026, the lowest per-token price on the board did not move at all: $0.06, then $0.06. Anyone tracking cost that way would report that the cost of intelligence had stopped falling. Over the same five months the lowest cost to actually answer the thousand questions fell from $0.625 to $0.100, a factor of 6.25.
One curve is flat and the other falls six-fold, over the same models, in the same window, from the same published prices. If you are deciding whether to build for a world where inference keeps getting cheaper, these two readings point in opposite directions.
Where the difference comes from
Entirely from output length. A reasoning model chooses how long to think, and that choice is invisible in a price sheet and enormous in a bill. Two models can be within a few percent on the rental rate and a factor of ten apart on what a task costs, purely because one of them writes three times as much before it commits to an answer.
This also means the rental rate is getting worse as a proxy over time, not better: as more of the field ships models that think before answering, the spread between rent and cost widens. The ratio in the table above goes 1.00, 4.13, 9.75, 14.12 across five months in one direction.
What to do instead
Price a task, not a token. Fix a workload you actually care about, run the candidate models on it, count the tokens they really emit, and multiply by their own rates. It is more work than reading a price sheet, and it is the only version of the number that survives contact with a reasoning model.
For self-hosted deployments the same principle applies with the terms swapped: what you are buying is machine-hours, so the question becomes how many tasks an hour of your machine completes, which is a utilisation question. And for agents, where the same context is read back a hundred times, the repetition dominates both.
The price frontier note plots twenty months of the board on the rental axis, because that is the axis the archive has for the whole window. This note is the caveat that belongs beside it: the horizontal axis there is rent, and rent is not cost.