Cost

Estimate spend before you make calls, and read the catalog's pricing. Both endpoints need only a valid key — no scope.

#Pricing catalog

GET /v1/pricing — per-model rates from the catalog: input/output cost per 1M tokens, context window, and max output tokens. It lists every enabled model in the catalog, unlike GET /v1/models, which is narrowed to the key's workspace and allow-list.

{ "object": "list", "data": [
  { "model": "gpt-4o", "input_cost_per_1m": 5.00, "output_cost_per_1m": 15.00, "context_window": 128000, "max_output_tokens": 16384 }
] }

#Simulate a cost

POST /v1/cost/simulate — estimate USD from token counts, or from sample text (a ~4-characters-per-token heuristic), optionally multiplied over many calls.

Field	Meaning
`model`	Public alias. Required; an unknown one returns `404 model_not_found`.
`promptTokens`, `completionTokens`	Token counts, when you know them.
`input`, `output`	Sample prompt / sample answer text, used to estimate the count a field above does not supply.
`requests`	Multiply the single-call cost by this many calls (default 1).

The two token-count fields are read in camelCase, as written above — the snake_case spelling other endpoints use is not recognised here and is silently ignored, which makes the estimate fall back to the sample text (or to zero).

{ "model": "gpt-4o", "promptTokens": 1200, "completionTokens": 400 }

{ "model": "gpt-4o", "input": "…a long prompt…", "output": "…a typical answer…", "requests": 1000 }

Response:

{
  "model": "gpt-4o",
  "prompt_tokens": 1200, "completion_tokens": 400, "estimated": false,
  "input_cost_usd": 0.006, "output_cost_usd": 0.006, "cost_per_call_usd": 0.012,
  "requests": 1, "total_cost_usd": 0.012
}

estimated is true when either count came from sample text. The simulator prices every prompt token at the full input rate; it does not model prompt caching or the cached-input rate.

#Where cost shows up

Non-streaming chat completions (including agent runs and /v1/responses) carry an X-Relay-Cost-Usd header — 0 on a cache hit or an idempotent replay. Streamed and embeddings calls record cost in telemetry only, because the usage is not known when the headers go out.
Every generation records its cost in telemetry, attributed per workspace / app (API key) / model / provider / end user / agent / prompt version. Cached prompt tokens are priced at the model's cached-input rate when one is set.
The panel's Cost page shows actual spend from telemetry and wraps /v1/cost/simulate; the Dashboard and Telemetry pages show real spend; Budgets enforce caps. See Telemetry & dashboards.