AI Model List

A public, curated JSON registry of AI model metadata — pricing, capabilities, context limits, and benchmarks. Designed to be consumed by any project that needs structured, up-to-date information about AI models across providers.

GoModel is one of the primary consumers, fetching this registry at startup to enrich its dynamically discovered models. But the registry is provider-agnostic and self-contained — any AI gateway, cost tracker, model selector, or dashboard can use it.

How It Works

The registry provides a single models.json file with metadata that upstream provider APIs don't expose directly: pricing, capabilities, parameter constraints, and benchmark rankings.

• Three-layer merge — provider defaults → model defaults → provider-specific overrides
• Human-readable — prices in USD per million tokens, clear field names, no scientific notation
• Sparse/additive — only include fields with values; absence = unknown
• Supplement, not gate — consumers decide how to handle models missing from the list

Quick Start

# Install validation and test dependencies
pip install jsonschema pytest

# Build the registry with LMArena + LiveBench rankings,
# and optional Artificial Analysis score rankings
export ARTIFICIAL_ANALYSIS_API_KEY=your_api_key_here
python scripts/build_registry.py --report-md tmp/build/report.md

# Validate the generated registry
python scripts/validate.py --models models.json --schema schema.json

File Structure

ai-model-list/
├── .github/workflows/
│   └── update-models.yml    # CI/CD entry point
├── models.json              # The registry — single output file
├── models.min.json          # Minified registry artifact
├── pipeline/                # Shared normalization, resolve, render, ranking logic
├── registry/curated/        # Curated provider policies and aliases
├── schema.json              # JSON Schema for validation
├── scripts/
│   ├── build_registry.py    # Fetch sources and build models.json
│   ├── fetch_sources.py     # Snapshot external sources
│   └── validate.py          # Validate models.json against schema.json
└── README.md

Schema Reference

Top-Level Structure

{
  "version": 1,
  "updated_at": "2026-02-19T12:00:00Z",
  "providers": { },
  "models": { },
  "provider_models": { }
}

Override / Merge Order

When resolving data for openai/gpt-4o:

1. Start with providers["openai"] — provider-level defaults
2. Overlay models["gpt-4o"] — model-level defaults (provider-agnostic)
3. Overlay provider_models["openai/gpt-4o"] — most specific overrides

Null fields in a more specific layer inherit from the layer above. Non-null fields override.

Providers

Keyed by provider slug (e.g., openai, anthropic, google).

Field	Type	Description
display_name	string	Human-readable name (required)
website	string/null	Provider website URL
docs_url	string/null	API documentation URL
pricing_url	string/null	Pricing page URL
status_url	string/null	Status page URL
api_type	enum	Protocol family: openai, anthropic, gemini, cohere, mistral, custom
default_base_url	string/null	Default API base URL
auth	object/null	Auth config: type (bearer/header/query_param), env_var, header
base_url_env	string/null	Env var for base URL override
supported_modes	array/null	Supported model modes
default_rate_limits	object/null	Default rpm, tpm, rpd

Models

Keyed by canonical model name (e.g., gpt-4o, claude-opus-4-6).

Field	Type	Description
display_name	string	Human-readable name (required)
description	string/null	Short description
owned_by	string/null	Originating company slug
family	string/null	Model family/series
release_date	date/null	ISO date of release
deprecation_date	date/null	ISO date of deprecation
tags	array/null	Curated tags (see Tags enum)
modes	array	Supported operational modes (required, see Mode enum)
modalities	object/null	input and output arrays of: text, image, audio, video
capabilities	object/null	Boolean flags — absent key = unsupported or unknown (see Capabilities enum)
context_window	int/null	Max input tokens
max_output_tokens	int/null	Max output tokens
pricing	object/null	Pricing in USD per million tokens (see Pricing)
parameters	object/null	Request parameter constraints
rankings	object/null	Benchmark scores (see Rankings)

Provider Models

Keyed by provider/model (e.g., openai/gpt-4o, bedrock/claude-opus-4-6).

Field	Type	Description
model_ref	string	References a key in models (required)
provider_model_id	string/null	Actual API model string (if different from canonical)
enabled	bool	Whether this mapping is active (required)
pricing	object/null	Override pricing (null = inherit from model)
context_window	int/null	Override context window
max_output_tokens	int/null	Override max output
rate_limits	object/null	Provider-specific rpm, tpm, rpd
endpoints	array/null	API endpoint paths
regions	array/null	Region availability (null = global)

Pricing Object

All monetary values in USD. Token-based prices are per million tokens.

Field	Type	Description
currency	"USD"	Always USD
input_per_mtok	number/null	Per million input tokens
output_per_mtok	number/null	Per million output tokens
cached_input_per_mtok	number/null	Per million cached input tokens
reasoning_output_per_mtok	number/null	Per million reasoning tokens (o1, etc.)
per_image	number/null	Per generated image
per_second_input	number/null	Per second audio/video input
per_second_output	number/null	Per second audio/video output
per_character_input	number/null	Per character (TTS)
per_request	number/null	Flat per request
per_page	number/null	Per page (OCR)
tiers	array/null	Context-length pricing tiers

Enums

Modes — operational types (array, at least one required):

chat, completion, embedding, image_generation, image_edit, video_generation, video_edit, audio_speech, audio_transcription, rerank, moderation, ocr, search, responses, code_interpreter

Capabilities (boolean, only include when true; absent = unsupported or unknown):

function_calling, parallel_function_calling, streaming, system_messages, vision, audio_input, audio_output, video_input, pdf_input, json_mode, structured_output, response_schema, reasoning, prompt_caching, web_search, computer_use, assistant_prefill, video_editing, image_input_embedding

Tags (curated, optional):

flagship, budget, preview, beta, deprecated, legacy, multimodal, reasoning, long_context, fast, open_weight, fine_tunable

Modalities: text, image, audio, video

Example

Minimal example showing one provider, one model, and one provider_model:

{
  "version": 1,
  "updated_at": "2026-02-19T12:00:00Z",
  "providers": {
    "openai": {
      "display_name": "OpenAI",
      "api_type": "openai",
      "default_base_url": "https://api.openai.com/v1"
    }
  },
  "models": {
    "gpt-4o-mini": {
      "display_name": "GPT-4o Mini",
      "modes": ["chat"],
      "context_window": 128000,
      "max_output_tokens": 16384,
      "pricing": {
        "currency": "USD",
        "input_per_mtok": 0.15,
        "output_per_mtok": 0.60
      }
    }
  },
  "provider_models": {
    "openai/gpt-4o-mini": {
      "model_ref": "gpt-4o-mini",
      "enabled": true,
      "provider_model_id": "gpt-4o-mini-2024-07-18",
      "endpoints": ["/v1/chat/completions"]
    }
  }
}

Scripts

validate.py

Validates models.json against schema.json (JSON Schema Draft 2020-12) and checks referential integrity (all model_ref and provider slugs resolve).

pip install jsonschema
python scripts/validate.py
python scripts/validate.py --models path/to/models.json --schema path/to/schema.json

build_registry.py

Fetches upstream source snapshots, resolves canonical models, applies rankings, and writes models.json plus models.min.json.

python scripts/build_registry.py
python scripts/build_registry.py --report-md tmp/build/report.md

Set ARTIFICIAL_ANALYSIS_API_KEY to include optional Artificial Analysis rankings during the build.

fetch_sources.py

Internal helper used by build_registry.py to snapshot upstream sources before normalization and merge.

Consuming the Registry

Fetch models.json via HTTP and resolve model data using the three-layer merge:

https://raw.githubusercontent.com/ENTERPILOT/ai-model-list/main/models.json

GoModel

GoModel is one of the primary consumers. It fetches the registry via the MODEL_LIST_URL environment variable:

export MODEL_LIST_URL=https://raw.githubusercontent.com/ENTERPILOT/ai-model-list/main/models.json

• Fetched on startup + hourly refresh
• Non-blocking, best-effort (failures don't prevent startup)
• Merged using the 3-layer override (provider → model → provider_model)
• Attached as ModelMetadata to matching registry entries

Contributing

Adding a new model

1. Add the model entry to models with at minimum display_name and modes
2. Add provider_models entries for each provider that serves it
3. Run python scripts/validate.py to verify

Adding a new provider

1. Add the provider entry to providers with display_name and api_type
2. Add provider_models entries linking existing models to this provider
3. Run python scripts/validate.py to verify

Updating registry data

Update the curated inputs or source adapters, run python scripts/build_registry.py --report-md tmp/build/report.md, then run python scripts/validate.py --models models.json --schema schema.json.