Models Hub
User GuideUser Console

Usage & Logs

Log fields explained, task logs, usage analytics and how to investigate unexpected costs.

Edit this page

The platform records full detail for every call. Once you understand this page you can answer "where did this money go" yourself.

Menu location: General → Usage Logs / Task Logs.

Usage logs

Each row is one API call. The core fields:

FieldDescription
TimeWhen the request happened
TokenWhich API key was used — the most useful field for attributing spend
ModelThe model actually called
GroupThe supplier route the request took
UsageInput and output tokens, plus cache read/write; per-call models show a count
CostThe amount actually charged
DurationTotal request time; streaming requests also record first-token latency
RetryWhether retries occurred and how many
Request IDUnique identifier. Always provide this when reporting an issue — support can locate the exact request with it
TypeSee log types below

Click any row to expand full detail, including a prompt preview, the complete error message, and previews of generated images, audio or video.

Log types

Filter by type at the top:

TypeMeaning
ConsumeNormal API charges — what you look at day to day
Top-upCredits added (payments, redemption codes)
RefundHeld amounts returned after a failed call
ErrorFailed requests, with the reason
ManageAdministrator actions on the account, such as manual adjustments
SystemSystem events

Seeing Refund entries is normal, not a problem: an estimate is held when a request starts, and the difference is returned when the call fails or uses less than estimated.

Filtering

Filter by time range (24 hours / 7 / 14 days / custom), key, model and type.

The fastest diagnostic path is: narrow the time range, filter type to Error, then expand the specific row for the full message.

Task logs

Asynchronous work such as image and video generation does not appear in the synchronous logs — check Task Logs instead:

FieldDescription
Task IDUnique task identifier
Submit timeWhen the task was created
Progress / statusQueued, running, succeeded, failed
Submit resultThe upstream response at submission
Failure reasonThe specific error when a task fails

Asynchronous tasks settle when they complete; failed tasks refund the held amount.

Usage analytics

Besides per-call logs there are several aggregate views:

  • Overview (General → Overview): balance, recent spend, daily call and cost trends, model mix.
  • Dashboard (General → Dashboard): finer breakdown by model and over time.
  • Cumulative spend per key: shown directly in the API Keys list.

Investigating cost

Costs higher than expected

Work through these in order:

  1. Filter logs by key — establish which key, and therefore which workload, is spending. A spike concentrated on one key often means misuse or a leak.
  2. Check the model — is it a higher-priced model? Prices are in the model catalog.
  3. Check input tokens — long context, whole documents in prompts, and accumulated conversation history push input volume far above expectations. This is the most common cause of overspend.
  4. Check retry counts — frequent retries mean an unstable upstream and amplify consumption.
  5. Check cache fields — cached portions are cheaper; misses are charged in full.

Failed calls

Rows of type Error record the exact reason. For insufficient quota, unavailable models, bad parameters, no available channel and similar, see Troubleshooting.

Reconciliation

For formal reconciliation, do not page through logs — use the monthly statement and daily reconciliation features on the Billing page, which export to CSV. See Billing & Reconciliation.

On this page