Skip to content

About

Value for money

Value for money shows whether a model costs more or less per task than its benchmark results predict. 50 is fair value. Above 50, the model is cheaper per task than its benchmark results predict. Below 50, it costs more.

Some benchmark sites record what each run actually cost in dollars, next to the score it reached. We compare each model's spend per task with what models at the same score typically spend. Then we combine those comparisons into one number from 0 to 100.

Last calculated 22 Nov 2026. A model's rating can change over time as new benchmark results arrive and prices move.

Why it moves

Three things feed the number. It does not separate them.

  • Price per token. What the model's provider charges for the text going in and coming out. Cheaper tokens make every task cheaper.
  • Tokens used per task. How much text the model reads and writes to finish a job. Long reasoning or many agent steps multiply the bill, so a model with cheap tokens can still be poor value. Model pages show the tokens used per task beside each benchmark score that has them, and benchmark pages add a column so you can compare models on one board.
  • Benchmark strength. The score the model reached on the task. A higher score sets a higher bar for what the model can reasonably cost.

How to read it

poor50 fairgood47.5 to 52.50100← costs more per taskcheaper per task →
50 is fair value. Scores from 47.5 to 52.5 count as fair, lower ones as poor and higher ones as good.
  • Above 52.5: good value. The model costs less per task than its benchmark results predict. The higher the number, the bigger the saving.
  • 47.5 to 52.5: fair value. The model costs per task about what models with its results typically cost.
  • Below 47.5: poor value. The model costs more per task than its benchmark results predict. That can still be worth paying for strengths the benchmarks miss.
  • Little data keeps it near 50. A model measured on only a few benchmarks is pulled toward fair value until more results arrive.

The tables show the number on a coloured tile. The band word appears when you hover or tap it, and on each model's page.

Which models are rated

Only models whose task costs have been measured on benchmark sites get a rating. Right now that is 376 of the 682 chat models with an llminfo IQ, about 55%. A dash means not measured, not poor value.

What it is not

  • Not the list price. Two models with the same price per token can land far apart here, because one uses many more tokens per task. For the price you pay per token, see the Input and Output price columns in the model tables.
  • Not a token count. The Yap score measures tokens per coding task on their own. It barely tracks value for money (the correlation is close to zero), because price per token varies far more between models than tokens per task. How the Yap score works
  • Not a quality ranking. A modest model that is cheap to run can score higher than a frontier model. For capability on its own, see llminfo IQ.
  • Not a recommendation for your workload. The number blends many benchmark tasks. Your prompts, your mix of input and output, and the reasoning setting you choose can all cost more or less than the benchmarks did. Check the per-benchmark columns for the work closest to yours.
  • Not a percentage. 100 is not twice as good as 50. Only the distance from the fair-value 50 carries meaning: above it the model is cheaper per task than its results predict, below it costlier.

How much data sits behind a rating

Select a score in the table to see how much data sits behind it. A rating with little data behind it is a rough estimate and is shown as one.

  • highPlenty of data behind it: eight or more benchmarks with measured task costs.
  • mediumA fair amount of data behind it: three to seven benchmarks with measured task costs.
  • lowLittle data behind it: one or two benchmarks with measured task costs, shown as a rough estimate.