We ran a simple experiment for a month, and the result is unambiguous enough that we're changing the product over it. We gave every registered agent a pile of free credits. Almost none of them were ever spent — while in the same window, tens of thousands of trial calls happened without anyone being asked to spend a thing.
The conclusion is hard to argue with: agents convert on doing the thing, not on holding a balance. So on September 1 we're ending the 500-free-credit grant and keeping the part that actually worked — a small number of trial calls per registered wallet.
| Signal | Value | What it says |
|---|---|---|
| Wallets granted free credits | 789 | plenty of signups took the offer |
| Free credits granted, total | 393,900 | a real giveaway, not a token one |
| Credits ever spent | 1,090 | 0.28% activation |
| Trial calls made | 42,053 | the part that worked |
| Unique identities that tried a call | 5,049 | demand is real and broad |
| Registered wallets | 790 | 15.6% of triers register |
| On-chain settlements | 90 (11.52 USDC) | paid demand is real, and tiny |
The two numbers that matter are side by side in the middle of that table: 393,900 granted, 1,090 spent. For every credit an agent actually used, we handed out 361 that sat there unused. That isn't a conversion problem — it's a category error.
An agent deciding to pay for an API call is not like a consumer deciding to spend a gift card. A consumer will eventually spend a balance because they're in the store anyway. An agent spends only when a specific call is the cheapest path to a specific result, right now. A dormant balance doesn't pull an agent toward the next call; it just sits there until the agent's own task logic produces a need.
Trial calls are different. A trial call is the product itself: the agent does the thing, gets the result, and learns whether the endpoint is worth paying for later. That's a decision an agent makes constantly. The giveaway that worked was "try it," not "have some."
This matches what the on-chain side shows: 90 real settlements, 11.52 USDC. Paid demand concentrates in the endpoints that do a concrete, repeated job — utility and audit work — not the long tail. A credit balance is a blunt instrument aimed at a market that pays per call.
Concretely, three things:
Registration still matters — it's what separates a trial identity from a paid one — but it no longer comes with a number that's going to quietly expire unused.
The lesson for the agent economy. Nobody is buying API access on credit. Agents pay for calls that do a job, one call at a time. If you're trying to grow a pay-per-call marketplace, spend your acquisition budget on making the first call effortless, not on making the balance bigger.
We'll watch whether anything changes on the other side of this — whether the trial-to-registration ratio moves, whether paid calls shift, whether anyone actually misses the credit grant. If the data says otherwise, we'll say so. That's the whole point of running it as an experiment instead of a policy.