modelscan registry

English | 简体中文

An open registry of large-language-model metadata. One machine-consumable JSON file — models.json — describing model identity, authorship, modalities, context/output limits, capabilities, lifecycle, and per-source commercial offers (prices, endpoints, rate limits) with the originating source kept as provenance.

Public site: https://modelscan.io/

What's covered

1,200+ models from every major lab, across every modality — text, image, video, audio, embeddings and rerank — each with identity, modalities, context / output limits, capabilities and pricing in both USD and CNY. A small sample of what's inside:

• Chat & reasoning — OpenAI GPT-5, GPT-5 Codex, GPT-4o, o-series; Anthropic Claude Opus, Claude Sonnet, Claude Haiku; Google Gemini; xAI Grok; DeepSeek V3 / R1; Qwen3; Meta Llama; Mistral / Codestral; Moonshot Kimi K2; Zhipu GLM-4.5; MiniMax M1 / M2.
• Image generation — Google Nano Banana (Gemini 2.5 Flash Image) and Nano Banana Pro, Imagen; OpenAI GPT Image; Black Forest Labs FLUX; ByteDance Seedream; Recraft; Kling Image.
• Video generation — OpenAI Sora 2; Google Veo 3; ByteDance Seedance 2; Kuaishou Kling; MiniMax Hailuo; Alibaba Wan.
• Embeddings & rerank — OpenAI text-embedding-3; BAAI BGE; E5; Gemini Embedding; Cohere Rerank; Voyage.
• Speech & audio — Whisper, Kokoro, and other TTS / transcription models.

Facts and prices are merged from OpenRouter, LiteLLM, Alibaba Bailian (百炼) and Volcengine Ark (火山方舟), so a single model can carry both USD and CNY pricing side by side.

Consume it

The canonical, always-current file:

https://raw.githubusercontent.com/modelscan/registry/main/models.json

curl -s https://raw.githubusercontent.com/modelscan/registry/main/models.json | jq '.models | length'

const { models } = await fetch(
  'https://raw.githubusercontent.com/modelscan/registry/main/models.json',
).then((r) => r.json())

The file is a single object: { schema_version, generated_at, count, models[] }. It is validated in CI against schema/models.schema.json (JSON Schema, draft 2020-12).

What a model looks like

{
  "id": "claude-opus-4-7",                  // canonical id (URL slug / program reference)
  "model": "Claude Opus 4.7",               // display name
  "author": "anthropic",                    // developer (a provider id)
  "alias_id": ["anthropic/claude-opus-4-7", "us.anthropic.claude-opus-4-7"],
  "input_modalities": ["text", "image"],
  "output_modalities": ["text"],
  "context_length": 200000,
  "max_output_tokens": 64000,
  "reasoning": true,
  "tool_calling": true,
  "release_timestamp": 1730000000,
  "endpoints": ["chat"],                    // API operations any source exposes
  "other_parameters": { "knowledge_cutoff": "2025-03" },
  "offers": [                               // one per source — prices, route, limits + provenance
    {
      "source": "openrouter",
      "currency": "USD",
      "prices": [{ "input": { "amount": 15, "unit": "per_1m_tokens" },
                   "output": { "amount": 75, "unit": "per_1m_tokens" } }]
    }
  ]
}

Key ideas

• Stable identity. Every model has one canonical id. Dated snapshots fold to their base id, and the dated / vendor-prefixed forms are preserved in alias_id — so the same model is never split into two rows across sources. author is always a provider id, so a developer never appears under two spellings.
• Two currencies. Pricing is kept in its native currency — never lossily converted: USD offers from OpenRouter / LiteLLM, CNY offers from Alibaba Bailian (百炼) / Volcengine Ark (火山方舟). A single model can carry both, side by side.
• Facts vs offers. Top-level fields are source-agnostic facts merged per field across sources. Commercial data (prices, endpoint paths, rate limits) lives in offers[], one per source, each carrying its source as provenance — so you can see where every number came from.
• Tiered & conditional pricing. prices[] is a list of tiers; a tier may carry conditions (input-length thresholds, or a variant label for axes like video resolution / audio).
• Lifecycle. A model that disappears from every source is marked deprecation: { status: "delisted", since } and kept, never deleted.

See schema/models.schema.json for the full contract.

Contributing

Corrections and additions are welcome — see CONTRIBUTING.md. models.json is machine-generated, so fixes are applied as maintainer overrides rather than direct edits to the generated file.

License

models.json, the schema, and the docs are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Use it anywhere, including commercially — just attribute modelscan registry (https://modelscan.io/).