Skip to main content
Model leaderboards usually rank benchmarks. A model report ranks something harder to fake and more useful to a builder: how much work developers actually put through each model, and what that work cost them. /metrics/models ranks up to 100 models by their trailing 30-day token volume, each with a sparkline of that window. Selecting one opens a full report at /metrics/models/{model} covering up to 365 days. The windows and ranking rules behind both are on How Usage Reports Work.

​
The directory

The directory lists ranked models with their provider, token volume over the ranking window, movement against the prior window, and publication status. It can be sorted, searched by name, and filtered by the company behind the model.

​
What a model report shows

  • Observed tokens and estimated spend, each with the change against the prior window.
  • Active developers, and how that count moved.
  • Estimated spend per 1M tracked tokens, which is the figure to compare across models rather than raw spend.
  • Cache-read share of the model’s traffic.
  • The coding client where the model sees the most use, with that client’s share of its tokens.
  • Momentum against the prior non-overlapping window, withheld rather than guessed when no prior window exists.
Spend per 1M tracked tokens is the comparison that survives across models. Total spend mostly tracks popularity, while the per-million figure reflects what a model costs to actually work in, and it moves with cache behavior as well as list price.

​
Where each model runs

Every model report ranks the coding clients the model was used inside, by share of its tokens. That distribution is the part list pricing and benchmarks both miss. A model can be dominant overall while being concentrated in a single harness, or spread evenly across many, and the two profiles imply very different things about how it is being adopted. Each entry links through to that client’s own report.

​
Reading a report honestly

These pages measure adoption, not quality. A model near the top is the one developers routed the most tokens through in the window, which reflects availability, default configuration, and price as much as capability. A model that is fast and cheap will out-token a model that is slow and careful even where the second does better work. The figures are also priced estimates drawn from tracked usage rather than billed amounts, and they cover only what was submitted. Read them as a picture of what developers ran, not as a verdict on which model is best.

​
Next steps