Call logs
Last updated September 15, 2026
The complete record of every call — status categories, routing attempts, itemized charges, artifacts, and how to locate a call by request ID.
The logs page records every call that passes through the gateway. It is the most direct place to answer "why did this fail" and "how was this charge calculated".
A record takes a few seconds to appear after a call completes. Not finding a request you just made is normal.
Locating a call
Every response carries an x-amux-generation-id header, of the form gen_…. Filter by request ID on the logs page to locate it precisely.
You can also filter by time range, model, provider, API key, status, or member.
List columns
| Column | Notes |
|---|---|
| Time | When the call was made |
| Model | The model actually routed to. If the request used an alias, "requested as …" is shown alongside |
| Provider | The one that served it |
| Protocol | The entry protocol used |
| Streaming | Whether the call was streamed |
| Status | See below |
| Duration / TTFT | Total time, and time to first token for streaming calls |
| Input / output / cache read / cache write | Token counts |
| Image output / video output / input media | Output quantities for media calls |
| Cost | The actual charge for this call |
| API key | The key used |
Status categories
| Status | Meaning | Charged |
|---|---|---|
| Success | Completed normally | Charged |
| Canceled | The client disconnected midway | Charged for usage already produced |
| Request error | Parameter errors, authentication failures, limit exceeded, and other client-attributed problems | Not charged |
| Upstream error | The provider returned an error or was unreachable | Not charged |
| Upstream truncated | The stream was cut off by the provider | Charged for usage already produced |
| Platform error | A problem on our side | Not charged |
Records with no charge are marked "not charged".
Detail view
Routing attempts
Lists the providers tried in order and the outcome of each. When failover occurred, this shows why the first choice failed and who ultimately served the call.
Failed attempts are not billed; the upstream cost is absorbed by the platform.
Failure reason and processing notes
On failure, the platform's error type is shown along with as much of the upstream response as possible — including details worth having when debugging, such as CUDA out of memory.
"Processing notes" record adjustments the platform made during the call, such as parameters dropped or a model name resolved through an alias.
Itemized charges
Lists each billing component with its quantity, unit price, and amount, using the same definitions as billing.
When a time-of-day price applied, the window is shown; promotions and account-specific discounts are shown separately. For the pricing rules, see billing overview.
⚠️ "Some usage did not match a price tier" means that portion was not charged. It usually indicates the provider returned a usage dimension the platform has not yet priced — please report it through a ticket.
When token counts are estimates
If a streaming call ends before the provider reports final usage — because the client disconnected early, for instance — token counts are estimated from the content transferred, and the detail page says so explicitly.
Media artifacts
Detail pages for image and video calls show the reference material supplied and the results generated, which can be opened or downloaded directly.
⚠️ Artifact links are cleared on a schedule; download anything you need to keep. Artifacts are stored in publicly readable object storage — the address is unguessable but requires no sign-in, so do not forward it to anyone who should not see the content. See data retention and privacy controls.
Raw JSON
The detail page offers the raw JSON of the record, convenient for pasting into a ticket or feeding into your own systems.
What is not recorded
The bodies of text and chat calls are not recorded. Logs contain metadata only — no prompts and no model output.
Prompts, parameters, and material references for image and video tasks are stored with the task and cleared after 30 days by default.
Visibility
| Role | Scope |
|---|---|
| Organization owner / admin | Organization-wide |
| Organization finance | Not visible |
| Workspace admin | All of their workspace |
| Workspace member | Their own only |
Reporting a problem call
The detail page offers a "report" action that opens a ticket with that call already attached, so there is no need to copy a request ID by hand. See getting support and tickets.
Related
- Usage and cost analytics
- Routing and failover — how failover is decided
- Errors and retries — what each error type means