Reference
The request log
One row per call, why it cost what it cost, and what is deliberately absent.
Every settled call writes a record. It carries who called, which name they asked for, what served it, how many tokens moved and what that cost on both sides. It carries no prompt and no completion.
seams requests --since 24h --limit 4
# ID AT END USER MODEL ACCOUNT STATUS TOKENS COST USD
# req_01JQZ8N4KP…0000 2026-09-07T09:20:00Z alice fast openai ok 1522 0.003044
# req_01JQZ8N4KP…0002 2026-09-07T09:20:00Z chen cheap openai capped 402 0.000804--since and --until take an ISO date or a span, and a command given neither covers the last 30 days.
| Status | Means |
|---|---|
ok | The provider answered and the call settled |
capped | A ceiling stopped it, mid-stream if it was streaming |
refused | It never reached a provider: no credit, or the model was not allowed |
upstream_error | The provider failed; the hold was released, not settled |
Why one call cost what it cost
explain is the answer to a support ticket. It shows the resolution, the token counts, what the call cost to serve and what it billed, so the arithmetic can be checked rather than trusted.
seams explain req_01JQZ8N4KP00000000000000
# requested model fast
# resolved model openai/gpt-4o-mini
# provider openai
# outcome support-reply
# status ok
# input tokens 1204
# output tokens 318
# provider cost micros 3044
# billed micros 3652
#
# pricing
# paid with balance
# markup bps 2000paid with says what the call drew on: an allowance the plan includes, the end user's balance, or the customer's own provider key. markup bps is what was billed over what it cost. Both come from the call's own record, so a disputed bill is settled by the row rather than by today's prices.
What is not in it
usage_record has no messages, no prompt, no completion and no body, and no jsonb column either. This is a property of the schema, not a setting: there is nothing to turn off, nothing to leak in a log line, and nothing for a subpoena to reach.