AtmoRouter crossed 250 accounts recently. That is small by gateway standards, but it is enough traffic to see real patterns instead of guesses — and the patterns are more interesting than the growth number.
The shape of usage
The distribution is what you would expect from agent-heavy workloads: a small number of keys produce most of the requests, while the majority of accounts use their key lightly — trying a model, running a small integration, testing an agent loop. Between those sits a middle band of steady, continuous callers. Nothing about this is unique to us; it is what API usage looks like everywhere. What is specific to AtmoRouter is which models that traffic chooses when every price is printed next to the official rate.
Price visibility changes the mix
When people can see that a strong general model serves at a fraction of the official list rate, behavior follows. Fast, cheap models carry the bulk of routine turns — classification, formatting, tool-call parsing, the inner loop of agent scaffolding — while the premium tier is reserved for moments that actually need it. That is the correct routing pattern, and making per-token rates legible per model seems to push people toward it faster.
Cadence tells you what people are building
The request rhythm is unmistakably agentic: bursts of short completions with steady streaming, concentrated in working hours, with long idle gaps in between. That is agent loop behavior, not chat behavior. It also explains the two features people lean on hardest: streaming on both protocols, and the never-bill-failed-requests policy — agent frameworks retry aggressively, and billing attempts would penalize exactly this traffic.
A free tier that is actually useful
One deep-research model is served at $0 per token (gated behind a small lifetime deposit cap to prevent abuse, not to nickel-and-dime). The pattern is consistent: people use the free model to validate their integration end-to-end, then move the same code to the paid tier by changing one string. That is the funnel working as designed — no trial timers, no crippled rate limits, just a model you can build against.
What we watch next
The interesting question as the catalogue grows is whether usage concentrates further on a handful of families or spreads as people optimize per-task. We will write it up when the answer is data instead of anecdote.