Get started

LiteLLM cost map

Source

Witness's attested language-model and retrieval rates rendered in the JSON dictionary format LiteLLM's --model_cost_map_url (LITELLM_MODEL_COST_MAP_URL) loads directly, with a vitna provenance block on every entry.

text
GET /api/v1/gateways/litellm/model-prices.json
bash
curl https://vitna.ai/api/v1/gateways/litellm/model-prices.json

Point a LiteLLM proxy at the live record without maintaining a hand-edited JSON fork:

bash
LITELLM_MODEL_COST_MAP_URL=https://vitna.ai/api/v1/gateways/litellm/model-prices.json litellm --config litellm.yaml

Response

json
{
  "openai/gpt-5.4-pro": {
    "litellm_provider": "openai",
    "mode": "chat",
    "input_cost_per_token": 0.00003,
    "output_cost_per_token": 0.00018,
    "input_cost_per_token_batches": 0.000015,
    "output_cost_per_token_batches": 0.00009,
    "vitna": {
      "sku": "gpt-5.4-pro",
      "providerId": "openai",
      "class": "model",
      "sourceUrl": "https://developers.openai.com/api/docs/pricing",
      "confirmedAt": null,
      "stale": false,
      "derived": false,
      "evidenceSha256": null,
      "citationUrl": null,
      "bands": [
        {
          "fromTokens": 0,
          "toTokens": 272000,
          "label": "to 272K tokens",
          "inputCostPerToken": 0.00003,
          "outputCostPerToken": 0.00018
        }
      ]
    }
  },
  "meta": {
    "generatedAt": "2026-09-12T09:14:02.881Z",
    "source": "database",
    "skuCount": 260,
    "keyCount": 590
  }
}

Keying and units

  • Per-token rates: Witness stores token meters per million tokens (input_per_mtok, output_per_mtok, embed_per_mtok) to six decimal places. This endpoint divides each figure by 1,000,000 rounded to twelve decimal places so all six digits survive without floating-point tail noise.
  • Aliases emitted: Each SKU is indexed under its Witness <slug>, <providerId>/<slug>, and <litellm_provider>/<apiModel> when the SKU register records a wire model identifier. First-party rows claim bare unprefixed names ahead of hosted resellers.
  • Context bands: Base-band rates populate the top-level fields. Exact 128K and 200K thresholds also populate LiteLLM's tiered keys, while the full band schedule is always preserved in vitna.bands so non-200K thresholds (such as 272K) are never misstated.
  • Provenance: When served from vitna.ai (meta.source: "database"), vitna.citationUrl resolves directly to /cite/{id} and vitna.evidenceSha256 names the archived pricing page.

Caching

public, max-age=60, stale-while-revalidate=3600.

Next step

Pair this cost map with local token profiling (vitna-runtime-profile) and CI eval normalization without changing your production gateway first: see Three-minute gateway quickstart and Gateway adapters.