Scope: InferenceView's retained transaction observations for September 12, 2026, 00:00–24:00 UTC. This combines Tempo RPC records with retained external-source observations. It is not a census of agent payments, customers or the market.
A one-day snapshot of InferenceView's public transaction data shows a sharp difference between the average and the middle of its known-value distribution. The average nominal value was $49.79, while the median was $0.002. Both figures describe the same 113,943 records with recognized nominal values, including known zeros.
The finding matters for anyone using payment dashboards to understand agent commerce: an average can give a very different impression from the typical observation in the same dataset. Neither statistic, by itself, establishes what an AI agent paid for a service.
The snapshot was retrieved on September 13 at 19:19:11 UTC, using a fixed, complete UTC calendar day. Its numbers describe the records retained at that retrieval time; later ingestion can change a historical result.
The snapshot in numbers
| Measure | September 12 UTC snapshot |
|---|---|
| Retained transaction observations | 125,711 |
| Records with recognized nominal values | 113,943 |
| Records with unknown nominal values | 11,768 |
| Unknown-value share of all records | 9.36% |
| Recognized nominal total | $5,673,526.391692 |
| Average among known-value records | $49.7926717 |
| Median among known-value records | $0.002000 |
Amounts use selected stablecoins at a nominal $1 per token. They are not measured market-dollar prices, merchant revenue or verified purchase values. Fees are separate.
Why the average needs a companion
The average divides the recognized nominal total by the number of records with known values. The median is the middle value after those same values are sorted. Unknown values enter neither calculation; known zeros enter both.
In this snapshot, a reader shown only the average might picture observations measured in tens of dollars. A reader shown the median would see a value below one cent. The difference reveals an uneven distribution in the recorded values, rather than a single representative payment size.
It does not reveal why the distribution has that shape. This dataset combines different sources and networks, and the aggregate does not establish buyer intent, the service purchased or whether a useful result was delivered. We therefore do not interpret the difference as evidence of agent spending habits or a market trend.
The missing 9.36% matters too
The API retained 11,768 records without a recognized nominal value. Those observations have not been converted to zero. A headline about the total without its known-value count would hide that boundary.
Unknown value can reflect unsupported tokens or evidence that cannot be reduced reliably to one recognized input. This snapshot does not provide a cause-by-cause breakdown of the unknown records. It also cannot tell us what their total economic value might be.
For researchers, the practical lesson is to report the observed count, the known-value denominator and the unknown count together. That makes changes in measurement coverage visible alongside changes in recorded activity.
Coverage is part of the result
The response reports a full 86,400 seconds of contiguous retained Tempo block coverage for the selected day. That coverage applies to Tempo. It does not certify complete coverage of the external networks included in the combined result.
The same response lists 41,405 retained external observations on Base, 18,486 on Solana and two on network eip155:4663. External data is source-indexed or sampled. Backfill was still underway for some sources at retrieval, so missing periods are not treated as zero activity and later queries may return additional records.
Records are keyed by transaction hash on Tempo and by network plus transaction hash for external observations. This avoids counting the same retained transaction again merely because another configured source reports it. It does not establish that each distinct transaction is a separate economic purchase. Conflicting or complex external transfers can be excluded by the ingestion rules.
Reproduce the analysis
Download the Python reproduction script. It can verify the saved snapshot offline or retrieve the same public query. A later query may include backfilled records.
python3 reproduce-transaction-values.py --read transaction-value-snapshot.jsonThe query uses inclusive since=1789171200 and exclusive until=1789257600. Readers can run the public aggregate query, download the retained reporting snapshot, and compare the API's nominal, priced_transactions and median_nominal fields. The historical snapshot preserves the exact result used here.
InferenceView publishes this analysis from its own observation service. For methods and limits, read how to measure agent payments and the API documentation. The useful next question is whether a result remains informative after its coverage, value assumptions and evidence limits are made explicit.