35 customers share one AI model (Claude Haiku). One customer, pebbleapp, floods it. pebbleapp asked for about 9,700 requests a second, while Haiku can handle about 3,000 a second for all customers combined. About half of all traffic in the test goes to this model from that one customer. Press play and watch what Turnstile does with everyone else's requests. This replay uses the round-robin policy: customers with waiting requests take turns, one request each.
Each box is a customer using this model. waiting = requests in their line right now, served = answered so far. A green outline flashes when they get an answer.
Tokens used in the last 60 seconds, against the model's limit. Turnstile stops sending when it reaches the dashed line.
Top: everyone else's requests answered each second. Bottom: pebbleapp's requests, answered (dark) and timed out after waiting too long (red). Click the chart to jump.
Replay of 150 seconds of Turnstile's benchmark (Apple M1 Pro, 640 simulated clients). The test uses a fake AI model that answers in 1–10 ms instead of the real 0.5–5 s, so it is about 500× faster than real life, and the token limit is scaled up by the same amount. Requests give up after 60 ms of test time (30 s in real terms). Counts are per second of test time.