LLM Rate Card

A daily-curated JSON artefact of per-model pricing and capability metadata for the LLM provider ecosystem. Built for TensorFoundry's products (notably Alloy, our LLM gateway) and published openly so anyone else can use it too.

Schema

v1.0.0 is published. Future versions live in sibling directories.

Version	Status	Schema	Example	Notes
v1	published	schema.json	example.json	v1/README.md

provider, mode, and capabilities are open snake_case patterns. Canonical known values live in schema/v1/registries.json and the pipeline cross-checks every emitted value against that file. See the authoring guide for how to edit the schema or cut a new version.

What's in each release

A single rate-card.json attached to a calver tag (yyyy.mm.dd, -N suffix for same-day reissues). Per-model fields:

• Input, output, cache read, cache write, and reasoning prices (per 1M tokens)
• modality_pricing for multi-modal models (per-modality rates for audio, image, video alongside the text default)
• Above-threshold pricing tiers for long-context models
• Context window, max output tokens, capability flags
• Provenance: which sources contributed, when a maintainer last verified the entry

A release is cut only when the canonical content actually changes. Automated runs that produce no diff do not tag.

Example entry

{
  "key": "openai:gpt-realtime-1.5",
  "provider": "openai",
  "model_id": "gpt-realtime-1.5",
  "mode": "chat",
  "input_per_million": 4.0,
  "output_per_million": 16.0,
  "cache_read_per_million": 0.4,
  "context_window": 32000,
  "max_output_tokens": 4096,
  "capabilities": ["streaming", "tools", "audio_input", "audio_output"],
  "modality_pricing": {
    "audio": { "input_per_million": 32.0, "output_per_million": 64.0, "cache_read_per_million": 0.4 },
    "image": { "input_per_million": 5.0, "cache_read_per_million": 0.5 }
  },
  "verified": "2026-05-03",
  "sources": ["litellm", "openai-direct"],
  "source_url": "https://developers.openai.com/api/docs/pricing"
}

Top-level fields (input_per_million etc.) carry the text rate. Non-text modalities live in modality_pricing. The sources array lists every contributor; source_url is the canonical pricing page.

How to consume

Grab the latest release artefact:

curl -L https://github.com/tensorfoundrylabs/llm-rate-card/releases/latest/download/rate-card.json -o rate-card.json

Look up a model by <provider>:<model_id>:

import json
card = json.load(open("rate-card.json"))
by_key = {m["key"]: m for m in card["models"]}
entry = by_key["anthropic:claude-sonnet-4-6"]
cost = (input_tokens * entry["input_per_million"] + output_tokens * entry["output_per_million"]) / 1_000_000

For multi-modal models, fall back to modality_pricing[modality] when present, else top-level (text).

Sources

LiteLLM is the primary source. Direct scrapers overlay where they have authority and divergence detection flags disagreements rather than silently overwriting. Currently scraping: Anthropic, OpenAI, Mistral, GLM, MiniMax, Groq, xAI, Cerebras, Qwen.

Pipeline

A daily GitHub Action fetches every enabled source, merges with divergence detection, validates against the schema and registries, and tags a release only when content changes.

Local development:

make install        # install deps
make ready          # fmt + lint + lint-yaml + type-check + test
make pipeline       # run against the fixture, write to .cache/
uv run rate-card --help

See src/rate_card/README.md for pipeline structure and how to add a source.

Contributing

PRs welcome for whitelist additions, scraper fixes, and schema feedback. For schema changes, read the authoring guide first.

Attribution

Primary data derived from LiteLLM (MIT). The artefact carries this attribution in its attribution field so downstream consumers inherit it.

Licence

MIT. See LICENSE.