Datasetpublic

kyisaiah47/tooldrift-model-rankings

ToolDrift: OpenRouter model usage rankings, captured daily One row per model per ranking window per capture: its rank, the tokens and requests behind that rank, and its share of the window. The series shows which models the market actually routes work to, day by day. Rows in this cut 44,369 One row is one model in one ranking window on one capture day Cut 2026-09-04 Refreshed Monthly, on the first of the month Measured by ToolDrift Method… See the full description on the dataset page: https://huggingface.co/datasets/kyisaiah47/tooldrift-model-rankings.

sourceHugging Facecc-by-4.0updated 8d agoView on Hugging Face

0likes55downloads

Dataset Card

ToolDrift: OpenRouter model usage rankings, captured daily

One row per model per ranking window per capture: its rank, the tokens and requests behind that rank, and its share of the window. The series shows which models the market actually routes work to, day by day.


Rows in this cut	44,369
One row is	one model in one ranking window on one capture day
Cut	2026-09-04
Refreshed	Monthly, on the first of the month
Measured by	ToolDrift
Method	https://toolproof.thecompound.tech/methodology
Licence	Creative Commons Attribution 4.0 International
Publisher	Compound Labs
Also on	Hugging Face · Zenodo

How it is measured

The public OpenRouter rankings are fetched on a schedule and stored as captured, one row per model per window. Nothing is smoothed, interpolated or revised after capture.

What a citer needs to know

—A gap in captured_on is a day the capture did not run. It is left as a gap rather than filled, because an interpolated row in a ranking series is indistinguishable from a measured one once it is cited.
—time_window is OpenRouter's own window label, passed through unchanged.

Files

File	Format	Size
`tooldrift-model-rankings-2026-09-04.csv`	CSV	4.9 MB
`tooldrift-model-rankings-2026-09-04.json`	JSON	11.0 MB

Every cut is also a GitHub release with the same files attached, so a citation can pin the exact edition it quoted.

Schema

Column	Type	Empty in sample
`captured_on`	string	0.0%
`time_window`	string	0.0%
`model_id`	string	0.0%
`permaslug`	string	0.0%
`rank`	number	0.0%
`total_tokens`	number	0.0%
`prompt_tokens`	number	0.0%
`completion_tokens`	number	0.0%
`requests`	number	0.0%
`share_pct`	number	0.0%

Cite it

Compound Labs (2026). ToolDrift: OpenRouter model usage rankings, captured daily. ToolDrift, https://tooldrift.thecompound.tech. Cut of 2026-09-04. Creative Commons Attribution 4.0 International (CC BY 4.0). https://github.com/kyisaiah47/compound-datasets/tree/main/tooldrift-model-rankings

bibtex

@dataset{compound_tooldrift_model_rankings_2026,
  title     = {ToolDrift: OpenRouter model usage rankings, captured daily},
  author    = {{Compound Labs}},
  year      = {2026},
  publisher = {Compound Labs},
  url       = {https://github.com/kyisaiah47/compound-datasets/tree/main/tooldrift-model-rankings},
  note      = {Cut of 2026-09-04. Measured by ToolDrift, https://tooldrift.thecompound.tech},
  license   = {CC-BY-4.0}
}

The whole collection has a DOI, 10.5281/zenodo.22844096, at https://doi.org/10.5281/zenodo.22844096. It resolves to the newest Zenodo version. To pin the exact deposit this cut belongs to, cite 10.5281/zenodo.22844097 instead.

Attribution is the licence condition, and it is the only one. Quote a figure with the publisher, the product that measured it and the cut date, and there is nothing else to ask.