Every model on tensor.news carries one published price. That number is not an average, and it is not a median. It is one aggregator's own figure — LiteLLM's, when LiteLLM has the model, and models.dev's otherwise. The other aggregators are consulted, and when they disagree the row is marked disputed.
The record computed that disagreement and then threw it away. It is now published.
Of the 124 price facts the record cross-checks, 72 are unanimous — every aggregator quotes the same price, and there is nothing to disclose. The other 52 carry a real disagreement, and there the published number is:
| Count | Share of the 52 | |
|---|---|---|
| Published price is the highest quote | 39 | 75% |
| Published price is the lowest quote | 12 | 23% |
| Published price is a middle quote | 1 | 2% |
Forty price facts carry a published price above the lowest quote on file. Across those, the median premium is 0% — most models are quoted the same by everyone — but the mean is 215.8% and the maximum is 19,819.3%.
The widest disagreements are not rounding:
| Model | Price | Highest quote | Lowest quote | Spread |
|---|---|---|---|---|
| Qwen3-235B-A22B | output | models.dev $2.80 | LiteLLM $0.60 | 366.7% |
| DeepSeek-R1 | output | models.dev $1.68 | LiteLLM $0.42 | 300.0% |
| Qwen3-235B-A22B | input | models.dev $0.70 | LiteLLM $0.20 | 250.0% |
| Qwen2.5-72B-Instruct | input | OpenRouter $0.36 | models.dev $0.11 | 227.3% |
| Qwen2.5-7B-Instruct | input | OpenRouter $0.10 | models.dev $0.04 | 150.0% |
| GLM-5.2 | input | models.dev $1.195 | OpenRouter $0.006 | 99.5% |
GLM-5.2 is the one to look at. Two aggregators price its input tokens at $1.195 and $0.006 — a ratio of nearly two hundred to one. The record publishes $1.195, because LiteLLM is the higher-priority source and the published price is always a source's own figure.
That is a defensible rule. It is also the rule that means a reader comparing providers is comparing against a number chosen for consistency, not for cheapness — and it is the highest quote three times out of four.
The record never averages prices across sources, and that is deliberate: an average of two quotes is a number nobody quoted. But "highest quote" and "the source we trust first" are different rules, and on this data they coincide 89.5% of the time.
Where the highest quote comes from, across all 124 facts:
| Source | Holds the highest quote | Supplies the published value |
|---|---|---|
| LiteLLM | 88 | 96 |
| models.dev | 28 | 28 |
| OpenRouter | 8 | 0 |
LiteLLM holds the highest quote on 88 price facts and is the published source on 96 — so on this record, LiteLLM is usually the expensive one. OpenRouter holds the highest quote on eight facts and is never the published source, because its terms of service constrain republication: it can mark a price disputed, never supply the number.
The rule is sound and the disclosure was not. A reader looking at the pricing table had no way to know that the number in the cell might be $0.60 while another aggregator listed $2.80 for the same tokens.
Every row of /models/pricing that has more than one quote now carries the spread, next to
the verification chip:
±367% where the sources disagree by a third or more,±15% where they disagree slightly, marked as corroborated,Twenty-five rows carry a chip: 39 price facts have a spread, spread across the models those facts belong to.
The model page says it in words. Where the sources disagree on a price, the pricing block now names each source's own figure and the gap between them, and states that the published price is the consensus rather than an average.
OpenRouter's values are never shown — only its name and the spread it implies — because its terms of service do not permit republication. Everything else on that page was already publicly listed by the source that published it.
A spread is a fact about the sources, not about the price. It says two aggregators disagree about what a model costs. It does not say which one is right, and it does not say what any provider will actually charge you at inference time — that depends on the endpoint, the cache hit rate, and the discount negotiated for volume.
The record does not model providers. It holds one published price per model and the disagreement behind it. What it can now say is exactly how much of the price is a cross-checked fact and how much is one aggregator's word.
The other half of the picture matters just as much. Seventy-two of the 124 price facts are
unanimous — LiteLLM, models.dev and OpenRouter all quote the same number, or all quote the
same two numbers when there are only two sources. For those, the cross-check confirms rather
than complicates, and the pricing table shows a corroborated chip and nothing else.
So the picture is not "the prices are unreliable". It is: most models are priced the same way everywhere, and a minority are priced radically differently, and before this week the record told you the first number without acknowledging the second.
And the honest caveat on the headline: this is not a story about wrong prices. It is a story about a number that was checked and reported without saying what the check found. The disagreement was always in the record — 124 price facts carried it, 52 of them with a real gap, all of it computed and discarded.