Tarseek API · Pre-production

Usage & token accounting

Measurements and request traces

New requests use measurement version 2. API completion time starts at HTTP API entry and includes preparation and prior provider attempts. Hosted chat timing starts when it enters gateway preparation, after any preceding chat assembly. It excludes the client network. Attempt latency starts at provider dispatch. Streaming TTFT means first useful text, reasoning, tool arguments, or media; role-only events are recorded separately as first-event time. Non-streaming TTFT is not measured.

Requests show the selected primary deployment, conservative cost estimates and exclusion reasons. The estimate uses input size and maximum output with a cold cache; recorded cost uses reported usage and catalog pricing. Successful v2 requests alone contribute to timing averages. Provider breakdowns attribute each attempt's tokens and estimated cost to its provider.

Sanitized access events cover validation and authentication rejections as well as stream outcomes. Authenticated preparation failures appear in Requests. Access events contain correlation IDs and status metadata, not prompts, responses or keys. Optional improvement samples remain separate and consent-controlled.

The public model page uses Tarseek observations from the last 30 minutes, with at least 100 observations per published group. Latency percentiles use nearest rank. Workload groups separate region, streaming, token sizes, cache and reasoning. Reliability includes failures without usage and measures observed success, not uptime. External benchmark references and registry price history are labeled separately.

The V1 chat completion usage object uses:

prompt_tokens for input tokens
completion_tokens including visible output
completion_tokens_details.reasoning_tokens
prompt_tokens_details.cached_tokens for prompt cache hits
total_tokens
cost — a USD estimate based on provider-reported usage and catalog pricing; it is not a Tarseek charge
cost_details.currency — currently USD
Typed image, audio, video, document, request, page, duration, character, and megapixel units where applicable

Implemented

Successful chat requests require provider usage and return normalized token counts plus the recorded cost estimate. A provider response without required usage fails asprovider_response_error. Streaming returns the final usage and cost chunk when stream_options.include_usage is true. Provider-reported cost is authoritative when supplied; otherwise Tarseek settles against the deployment pricing-rule snapshot and measured usage. Your provider bills the connected BYOK account directly, and its invoice remains authoritative. Ambiguous or missing non-token measurements fail closed.