About
Yap score
Tokens used per coding task vs. models with the same results.
The Yap score shows how many tokens a model generates per task on coding and agent work. On each benchmark, we compare the tokens a model used per task with models that reached the same score there. Then we combine those comparisons across the coding and agent benchmarks into a number from 0 to 100. 50 is typical. Above 50, the model generates more tokens than models with the same results. Below 50, it generates fewer. It covers coding and agent work because that is one common place where token costs add up.
Coding here means agents that edit code and run commands, plus single-shot programming contests such as ALE-Bench. Contests agree only partly with agent coding tasks. Many models have only been measured on contests, so their score rests on one benchmark and is shown as a rough estimate.
Last calculated 22 Nov 2026. A model's score can change as new benchmark results arrive.
Why one score, and why coding
We first tried one Yap score across every benchmark that publishes token counts. It didn't hold up. How much a model talks depends on the kind of task. A model that's brief on maths problems can be one of the wordiest on coding tasks, and the reverse. Across the models we could check both ways, the two rankings had almost nothing in common. A blended number would describe neither.
Nearly all the token counts published today come from two kinds of benchmark: competition maths, and coding (including agents that edit code and run commands). So those were the two groups we could measure properly.
Maths verbosity is interesting, but it doesn't help you pick a model. It doesn't predict how a model behaves on coding or agent work, which is one common place where token costs add up. So the Yap score covers coding. If you want maths token figures, each maths benchmark page lists tokens per task, MathArena for example.
Some sources publish only a total that includes the text fed into the model. Those totals mostly reflect how each test setup manages context, not the model, so we leave them out.
We also checked agent tasks that aren't coding, like web research and customer service. So far they only loosely match coding, and there's too little data to score them on their own. We'll look again as more results come in.
Effort changes it
The same model generates more tokens at a higher effort setting, so every score names the effort it was measured at. On 27 Sep 2026, we had 128 cases of one model measured at two effort settings. The higher setting scored higher in 120 of them. Model pages show the score at each effort we measured.
A model with reasoning off is not automatically below 50. It is compared with models that reached the same results, so it can still generate more tokens than they do.
How to read it
- Above 55: Chatty. The model generates more tokens per task than models with the same results.
- 45 to 55: Typical. The model generates about as many tokens per task as models with the same results.
- Below 45: Terse. The model generates fewer tokens per task than models with the same results.
- The ± is the spread. The ± shows how much the score varies across the coding benchmarks the model was measured on. It doesn't include run-to-run noise on a single benchmark, which sources don't publish. Two models whose ranges overlap are not clearly separated.
Table colours show token use. Green means Terse, yellow means Typical, and red means Chatty. Scores are rounded to whole numbers before assigning a band.
The tables show the number and its ±. The band word and the effort appear when you hover or tap it, and on each model's page.
What it is not
- Not a quality score. It says nothing about how good the answers are. For capability on its own, see llminfo IQ and Coding IQ. How llminfo IQ works How Coding IQ works
- Not a price. It counts tokens, not dollars. For the price per token, see the Input and Output price columns in the model tables.
- Not value for money. 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 Value for money works
- Not a percentage. 100 is not twice 50. Only the distance from the typical 50 carries meaning.
Which models are scored
Only models whose token use has been measured on coding benchmarks get a score. Right now 151 models have a Yap score. Of those, 34 have a fair amount of data behind them and 117 little. A dash means not measured, never terse.
- highPlenty of data behind it: 8 or more groups of related benchmarks.
- mediumA fair amount of data behind it: 3 to 7 groups of related benchmarks.
- lowLittle data behind it: 1 or 2 groups of related benchmarks, shown muted as a rough estimate.