model-specs

A community catalog of LLM model specifications — pricing, context windows, capabilities, and client/CLI compatibility — kept fresh by automatically syncing LiteLLM's model database and layering objective corrections plus data LiteLLM doesn't track (which models each AI CLI actually accepts).

If you build AI applications and keep re-implementing the same "what does this model cost / what's its context window / does this CLI accept this model id" lookups, this repo is meant to be the single source you vendor or fetch.

What's in here

Directory	What it holds	Edit by hand?
upstream/litellm/	A read-only mirror of LiteLLM's model_prices_and_context_window.json, refreshed by CI.	No — auto-synced
patches/	Objective corrections to the upstream data (wrong pricing, provider/platform mapping, capability flags).	Yes — via PR
clients/	Which model ids each AI CLI / client accepts (Claude Code, Codex, Antigravity). LiteLLM does not track this.	Yes — via PR
dist/	The merged, validated artifact consumers read: catalog.json + catalog.sha.	No — generated
schema/	JSON Schema for dist/catalog.json.	Yes
scripts/	sync (pull upstream), build (merge → dist), validate (schema + invariants).	Yes

Design rule: upstream/ is never hand-edited. Every correction lives in patches/ so git diff always shows exactly what we changed vs. what LiteLLM shipped, and the merge stays reproducible.

What's intentionally NOT here

• Pricing markup / margin. This repo holds objective base costs only. Your selling price is your business policy — keep it in your own app.
• Which models a given product chooses to sell. That's product config, not a community fact. Keep served-model lists in your own project.

Consuming the catalog

Read dist/catalog.json. Two patterns:

1. Vendor + embed — add this repo as a git submodule and embed dist/catalog.json at build time. Reproducible, offline-safe, but updates need a rebuild.
2. Embed floor + scheduled refresh (recommended for services) — embed dist/catalog.json as the offline fallback, and have a scheduled job fetch the latest artifact (raw URL or a GitHub Release asset), verify catalog.sha, and upsert into your store. New models / price changes land without redeploying.

Shape of dist/catalog.json

{
  "version": "<sha256[:12] of the models+clients payload>",
  "generated_at": "<ISO-8601 — reflects upstream fetch time, not build time>",
  "upstream": { "source": "litellm", "sha": "<commit>", "fetched_at": "<ISO-8601>" },
  "models": {
    "gpt-4o": {
      "provider": "openai",
      "platform": "openai",
      "mode": "chat",
      "input_cost_per_token": 2.5e-6,
      "output_cost_per_token": 1.0e-5,
      "cache_read_input_token_cost": 1.25e-6,
      "cache_creation_input_token_cost": 2.5e-6,
      "cache_creation_input_token_cost_above_1hr": 5e-6,  // Anthropic extended cache
      "input_cost_per_audio_token": 1e-6,                 // multimodal models
      "output_cost_per_reasoning_token": 2.5e-6,          // reasoning models
      "max_input_tokens": 128000,
      "max_output_tokens": 16384,
      "capabilities": { "function_calling": true, "vision": true, "prompt_caching": true },
      "source": "litellm"
    }
  },
  "clients": {
    "claude-code": { "description": "...", "models": ["claude-sonnet-4-6"], "updated": "<date>" }
  }
}

All cost fields are USD per token (not per million). source on each model is litellm (verbatim from upstream) or patch (corrected here). platform is the normalized vendor (e.g. all vertex_ai-anthropic* → anthropic), distinct from LiteLLM's finer-grained provider.

Contributing

npm install   # first time only — installs ajv for validate

1. To fix a price/capability/provider: edit the right file under patches/.
2. To update a CLI's accepted models: edit clients/<client>.json.
3. Run npm run build && npm run validate and open a PR. CI rebuilds dist/ and fails on schema or invariant violations.

License

MIT. The data under upstream/litellm/ is mirrored from BerriAI/litellm (also MIT) — attribution retained in upstream/litellm/SOURCE_SHA.