Methodology

aimodel.directory is a reference, not a ranking. Our goal is that every number on the site can be traced to a primary or well-maintained public source, carries the date we last checked it, and is never changed without a dated record of the change.

Sources

  • models.dev (MIT licence) for specifications and first-party pricing. Each provider listing is matched to a cross-provider model identity, so the same model sold by its lab, a cloud and an inference host appears as one model with several price rows.
  • OpenRouter's public models API for gateway pricing, context lengths and published retirement dates, and for models models.dev does not yet list.
  • The Hugging Face Hub for open-weight licence, parameter count and adoption. We accept Hub data only when the repository name matches the model exactly; near-misses such as a fine-tuned or RL variant go to an editor instead.

The full list, with licences and attribution, is on the sources page.

When sources disagree

Each field has a ranked list of sources. A lab's own listing outranks a reseller's; the Hugging Face repository outranks API catalogues for licence, weights and parameter count; an editor's verified correction outranks everything. A lower-ranked source can fill a gap but never overwrite a higher-ranked value, so figures do not flip between daily runs.

Freshness and change tracking

Every source is re-read daily. We compare each run with the previous one and record every user-visible difference (a price move, a new provider listing, a context window change, a retirement date) as a dated event. Those events power the changelog and each model's change history, and they are the only thing that updates a page's modified date. Download counts change daily and are not treated as changes.

What we publish, and what we hold back

  • Prices are USD per million tokens, as listed by each provider. “Last verified” is when we last read the source.
  • A model page is indexed by search engines only when it has enough sourced facts to be useful and the model is attributable to a known lab or a live API listing.
  • Benchmark scores, when added, will be labelled as self-reported (published by the lab) or independent, with a link to the original result.
  • Summaries are generated from the data itself. Numbers are never written by a language model.

Open data

Every model page has machine-readable JSON and Markdown versions (add .json or .md to its URL), published under CC BY 4.0. See llms.txt for a guide for AI systems.