Machine-readable catalogue

Public dataset

ModelCap publishes a bounded language-model snapshot for reuse. The JSON and CSV files below are the same public projection: canonical identities, ModelCap ranks and evidence states, listed token prices, context limits, and current provider coverage. Media catalogues and live discovery feeds are not in this download. A separate evidence file and notebook let you rebuild every published score and board position yourself.

Snapshot generated 28 Sept 2026, 02:54 UTC · schema modelcap-public-models-v1

Downloads

Both files are CORS-open and refresh with the board snapshot. Use them as a catalogue, not as a live price API: a listed price is the OpenRouter observation sealed into this snapshot.

/data/models.json — JSON projection
/data/models.csv — RFC 4180 CSV of the same rows

curl -sS https://modelcap.ai/data/models.json -o modelcap-models.json
curl -sS https://modelcap.ai/data/models.csv -o modelcap-models.csv

Refresh cadence

The AWS data plane publishes a new sealed snapshot when sources change. This page, the sitemap lastmod values, and the downloadable files all use that snapshot clock. They never substitute the time of your request. HTML revalidates about every 60 seconds so crawlers and browsers pick up a newly published snapshot without a code deploy.

Rank method modelcap-index-v4.12-fused-launch-placement. Evidence admission modelcap-score-v10-best-supported-configuration. Scoring rules are documented on /methodology.

Field schema

The JSON object carries snapshot provenance, then a models array. CSV repeats the generation instant and method ids on every row.

JSON	CSV	Meaning
`modelUrl`	`model_url`	Canonical ModelCap model page for this identity.
`modelId`	`model_id`	Stable catalogue identity used internally and in exports.
`slug`	`slug`	URL slug for /model/{slug}.
`name`	`name`	Publisher-facing model name.
`publisher`	`publisher`	Publisher display name.
`category`	`category`	Always language in this export. Other media stay on separate catalogues.
`isCurrent`	`is_current`	False when a newer family member has superseded this identity.
`rank.current`	`current_rank`	Public current-board rank, or null when unranked.
`rank.allVersions`	`all_versions_rank`	Public all-versions archive rank, or null when unranked.
`modelCap.score`	`modelcap_score`	ModelCap Index score on a 0–100 scale, only when a public rank exists.
`modelCap.evidenceState`	`evidence_state`	The published evidence label, such as Measured, Measured · blended, Modeled · peer benchmarks, Modeled · succession or Modeled · lineage. Null when no public index is published.
`modelCap.confidence`	`modelcap_confidence`	Evidence-support percentage behind the published score.
`modelCap.interval.lower / upper`	`score_interval_lower / score_interval_upper`	Published score interval bounds.
`weightAccess`	`weight_access`	none, open, gated, restricted, or unknown.
`releasedAt`	`released_at`	Catalogue or repository listed release instant, as ISO-8601 UTC.
`contextTokens`	`context_tokens`	Published maximum context length in tokens. Null when unknown.
`pricingUsdPerMillionTokens.input / output`	`input_price_usd_per_million_tokens / output_price_usd_per_million_tokens`	Listed OpenRouter token prices in USD per 1M tokens. Null unless a fresh endpoint observation backs them.
`currentProviderEndpoints`	`current_provider_endpoints`	Count of currently observed OpenRouter serving endpoints.
`huggingFaceId`	`hugging_face_id`	Pinned Hugging Face repository id when one is published.

Reproduce the board

Every published ModelCap Index score and both board orders can be rebuilt from two files. The evidence file lists, for each ranked model, the inputs its score combines and the keys its position sorts on. The notebook rebuilds the board from that file using only the Python standard library.

/data/evidence.json — per-model inputs, refreshed with each snapshot
/data/modelcap-recompute.ipynb — Jupyter notebook that checks every score and position

A measured score starts from ModelCap’s 0–100 score for each benchmark row. Rows are weighted by their board’s share and their confidence, then averaged within each family. The general family then moves by each specialist family’s capped difference from it. A score without a general-family board is the precision-weighted mean of the estimates listed with it. Both boards then sort by score, basis, confidence and identity key. Published inputs carry one decimal, so a rebuilt score counts as matching when it lands within 0.15 points.

Two steps can’t be rebuilt from these files. Turning a raw leaderboard result into a 0–100 score reads each full upstream board, which ModelCap doesn’t redistribute; every row links to its source instead. The modeled estimates, such as verified lineage and launch-card placement, are fitted by the methods described on /methodology, and launch accuracy tracks how well they hold up.

curl -sS https://modelcap.ai/data/modelcap-recompute.ipynb -o modelcap-recompute.ipynb
jupyter notebook modelcap-recompute.ipynb

Licence and citation

ModelCap’s compilation — ranks, evidence states, confidence intervals, catalogue classification, and this documentation — is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Attribution requires the name ModelCap and a link to https://modelcap.ai/data or to the specific model page you used.

Underlying source facts remain subject to their publishers’ terms. OpenRouter listed prices, Hugging Face repository metadata, Arena and other evaluation rows, Vercel AI Gateway shares, and Polymarket contracts are not re-licensed by this page. If you redistribute those values, follow the original venue.

The export is provided as-is, without warranty that a price, context window, or rank remains current after the snapshot instant. Quote the generation timestamp with any figure you reuse.

Suggested citation: ModelCap. Language model rankings dataset (modelcap-public-models-v1). Generated 28 Sept 2026, 02:54 UTC. https://modelcap.ai/data

BibTeX:

@misc{modelcap_index,
  author = {{ModelCap}},
  title = {Language model rankings dataset},
  year = {2026},
  url = {https://modelcap.ai/data},
  note = {Snapshot 2026-09-28T02:54:15.503Z, rank method modelcap-index-v4.12-fused-launch-placement}
}