2026-11-09 · data notes
A free tier is usually argued about as a number: how many calls, for how long, per what. We have been running one on a pay-per-call x402 catalog for three months, and the ledger says the number is not the interesting variable. Who the free tier counts decides what it measures — and the same quota can produce two completely different curves depending on whether the identity behind it is an IP address or a signed wallet.
Numbers below are read from our own trial tables on 2026-11-09. They are small, they are ours, and they are reported as measured rather than as advice.
The trial ledger holds 10,497 global rows plus a legacy per-endpoint table that stopped receiving writes on 2026-08-11. Across those tables the same idea — "a few free calls" — ran under three different identity rules:
| regime | identity | window | identities | at cap |
|---|---|---|---|---|
| per-endpoint | IP or wallet, one counter per service | to 2026-08-11 | 2,350 rows | — |
| per-IP | source address | 08-13 → 09-01 | 4,634 | 3,247 with ≥5 calls (70.1%) |
| signed wallet | EIP-191 signature over a per-call message | 08-14 → 11-09 | 3,513 | 502 with ≥5 calls (14.3%) |
The per-IP regime looks, at first glance, like a system with enormous demand: seven in ten identities used five or more calls. It was not demand. Addresses are not people and not agents; they are the cheapest thing to mint, and the counter was measuring a pool of traffic with no per-caller continuity. Hitting a high count was the expected outcome, not a signal.
The signed-wallet regime reverses the distribution:
| calls used by a signed identity | identities | share |
|---|---|---|
| 0 | 424 | 12.1% |
| 1 | 2,448 | 69.7% |
| 2–4 | 139 | 4.0% |
| 5 or more | 502 | 14.3% |
(Four of the last row predate the current five-call cap; the other 498 stopped exactly at five.) Seven in ten signed identities use exactly one call and stop. Only one in seven exhausts the allowance. Whatever a free tier is for, on this surface it is a sample, not an evaluation budget: the caller is checking whether the shape of the answer is useful, and one answer is usually enough to decide.
1. A free tier is an identity mechanism before it is a quota. The quota we argued about was five calls. Under per-IP counting, those five calls were consumed by traffic that had no reason to continue; under signed counting, most callers never reach them. If you are choosing a number before you are choosing an identity, you are tuning a counter that is not measuring what you think it is.
2. If most callers stop after one answer, the first answer carries the whole product. A free call that returns a plausibly-shaped but incomplete result teaches the caller nothing it can act on. On the other side, a first answer that is complete enough to decide with turns the free call into the top of a funnel instead of a cost center. The trial budget is not where the conversion decision happens; the first response is.
3. Breadth does not create trial demand. Our catalog lists 1,856 services; the legacy per-endpoint table holds rows for 256 distinct service ids, and its top ten endpoints account for 964 of 2,350 identity rows (41%). The long tail is not under-sampled because the quota is too small — it is unvisited. Adding more endpoints to a catalog does not manufacture the intent to try them.
4. Anti-abuse and conversion pull in different directions, and you should know which one you are buying. Signing every trial costs the caller a keypair and a signature, and it removes the cheapest abuse vector. It does not make anyone pay. We would make the same trade again, but it should be described as what it is: an identity decision that changes the measurement, not a pricing decision that changes revenue.
When a free tier is measured through an identity that is cheap to mint, the number you get describes the minting cost. Pick the identity first; the quota only means something after that.
Concretely, the questions we would ask before setting a free-tier number on any machine-facing API: can this identity be created for free in under a second? does it persist across processes and days? can it be tied to the thing that will eventually pay? If the answers are yes, yes, yes — then a per-identity budget is a real measurement. If the identity is an address, a user agent, or a header the caller chooses, the budget is a limit on a traffic class, and it will be full, and that fullness will mean nothing.
Data source: minia2a's own trial ledger, read 2026-11-09. Global quota on this site is five trial calls per signed wallet, shared across the catalog.