We gave away 133 free API keys. 80 of them were never used.

On 6 September we put $10 of free credit on the table for 100 accounts, locked to one model, no card required. It filled up in a week and we closed it on 13 September. It also produced a clean dataset on how coding agents actually consume tokens. This is what the logs say happened, including the parts that did not go the way we expected.

We issued $1,330 of credit. People spent $155.80 of it.

133 grants went out in total, which is more than the 100 we planned, because suspended and revoked accounts release their slot back into the pool. That is $1,330 of theoretical credit. Actual usage across the whole life of the offer was $155.80, about 12% of what was issued, every request billed at its exact metered cost so anyone can check the charges against their own token counts.

The distribution is the part worth staring at:

• 80 accounts claimed a key and never sent a single request.
• 32 spent less than $1.
• 18 spent between $1 and $5.
• 3 spent between $5 and $10.
• Nobody exhausted their $10.

60% of the people who went through email verification, clicked an activation link and received a working API key then did nothing with it. Not "tried it and churned". Zero requests, ever. If you run a free tier and measure signups, that number is the one your dashboard is hiding from you.

96% of all input tokens were cache hits

The 58 accounts that did use it sent 22,165 requests and 2.5 billion input tokens. Of those input tokens, 2,495 million were cache hits and 30 million were not. That is 96.1%, and for the heaviest individual users it ran between 95% and 99%.

This is the whole argument for pricing agent traffic on the cache-hit rate rather than the headline input price, and it is nice to see it hold at this scale on traffic we did not curate. An agent re-sends its conversation on every tool call, so the same prefix is billed again and again. Whatever a provider charges for a cached token is, for practical purposes, what it charges you.

It also means the free tier stretched much further than $10 of list price suggests. The users who treated it as a real workload got a fortnight of coding-agent use out of it.

Cost lands in a handful of hours, not evenly

We bucketed every request into five hour windows per account. There were 279 such windows across the life of the offer. The five most expensive of them accounted for 45.6% of all spend. Over half of all windows cost less than fifty cents between them.

Agent usage is not a steady drip. It is someone opening a large repository at two in the afternoon and sending 1,800 requests in five hours, then nothing for three days. Any limit designed around requests per minute will miss this completely, because a single 800,000 token call and a 2,000 token call are one request each and differ in cost by two orders of magnitude. Daily and per window spend ceilings are what actually bound it, and the API documentation lists the limits and the 429 codes that report them.

What we would keep

The spend ceilings earned their place. A rolling daily cap per account, and cost based rather than request based limits, are the only reason a 2.5 billion token giveaway stayed predictable. Every account that hit a limit hit it with headroom left, and nothing ran away overnight.

The manual review queue also earned its place, though it is a real cost in attention. Custom domains, datacentre and VPN addresses and repeat attempts from one network were where essentially all of the bad signups came from. Personal addresses on consumer connections were, almost without exception, real people.

What we would change is the measurement. We tracked slots claimed, because that is what the cap counts. Slots claimed turned out to be a nearly meaningless number: 80 of them were people who never made a request. Active accounts and spend against the ceiling would have told us within two days what took three weeks to see.

The free tier is closed for now. Credit packs start at $10, never expire, and unlock every model in the catalogue on one key.