Pricing & usage

Per-token prices, per model, listed in the catalog and charged against your AhuraSense balance as requests complete. No subscription and no minimum.

How a request is priced#

Each model in the catalog has three rates, in cents per million tokens:

Rate	Applies to
`input_cents_per_mtok`	Tokens in the prompt: every message you send, tool definitions and tool results included.
`cached_cents_per_mtok`	Prompt tokens the backend served from its prompt cache, and the whole prompt of a response answered by the gateway cache.
`output_cents_per_mtok`	Tokens the model generated, reasoning tokens included on models that reason.

example

Model:  anthropic/claude-haiku-4.5   input $1.00 / M   cached $0.10 / M   output $5.00 / M
Usage:  prompt 1,200 tokens (of which 800 cached), completion 300 tokens

cost = 400 × 1.00/1M + 800 × 0.10/1M + 300 × 5.00/1M
     = $0.0004 + $0.00008 + $0.0015
     = $0.00198, recorded as 1 cent (costs round up to the cent per request)

Prices are on the Models page and in GET /v1/models. A change applies to requests after it, never retroactively.
Requests that end in an error are not charged. A cancelled stream is charged for the tokens generated.
The off_peak field on a model carries a time-of-day discount window; it is null on every model today. When one is set, prices.discount on the same model gives the discounted rates and says whether the window is open right now.

What a usage record contains#

Every request produces one record, written a few seconds after the response completes so that metering never adds latency:

Field	Meaning
`request_id`	Matches `X-Ahura-Request-Id`.
`model_id`	The catalog id that served the request.
`api_key`	Which key made it, shown by name and masked prefix.
`input_tokens, output_tokens, cached_tokens`	The counts the charge was computed from.
`cost_cents`	What was charged.
`status`	`success`, or an error class such as `error_rate_limit` or `error_upstream`.
`latency_ms`	Time from request receipt to the last byte of the response.
`billed_to`	How the request was charged: `platform`.
`cache_kind`	`none` or `l1`.

Prompts and completions are not part of the record. See Privacy & data retention.

Inference → Usage shows month-to-date spend, requests, tokens and latency percentiles, broken down by model and by key, with cache hit rates, and exports the raw records as CSV for any window up to 90 days. GET /v1/key gives a service its own month-to-date spend without the dashboard.

Something missing or wrong on this page? Tell us, and quote the page title.

How a request is priced#

What a usage record contains#

Where to see it#