Performance methodology
How to read
the results.
A trace-based comparison of model-call speed and inference cost, with separate estimates for complete workflows and annual impact.
The workload behind the replay
Applied Compute’s production workload profiles capture multi-turn agent sessions. A trace includes the session’s model requests and tool calls; its office-work profile covers documents, spreadsheets, and slide decks over a large filesystem with a substantial tool harness. Replay recreates the request sizes, growing context, and tool waits, rather than proving that a business task was completed successfully.
The Lab uses a selected office-work trace in a 57-model-call replay. Infaio results are modeled from measured hardware curves. OpenRouter figures are independent medians from billing-complete live replays: 15 speed-priority runs and 13 cost-priority runs. Model/provider choices differ, so the comparison does not isolate the router alone or establish equivalent task quality.
| Metric | Infaio Cost | OpenRouter Cheapest | Infaio Speed | OpenRouter Fastest |
|---|---|---|---|---|
| Inference cost | $0.129 | $0.410 | $0.195 | $1.093 |
| Model-call time | 21.62 s | 108.25 s | 17.14 s | 95.05 s |
| Tool allowance | 107.433 s | 107.433 s | 107.433 s | 107.433 s |
| Full-workflow estimate | 129.05 s | 215.69 s | 124.57 s | 202.48 s |
Model-call time includes prefill, generation, and call overhead; it excludes tool execution and client retries. Full-workflow estimates add the same 107.433-second trace-derived tool allowance to both paths. They are not measured end-to-end executions and include no additional Metro Edge or tool-colocation saving.
Annual cost and completions
At a fixed workload volume, annual inference cost = runs × cost per replay. Savings are the difference between the two paths. The separate same-time scenario estimates Infaio completions as floor(baseline runs × baseline full-workflow time ÷ Infaio full-workflow time), assuming unchanged concurrency, sufficient demand, and no other bottleneck.
The scenarios use full-precision values and exclude other operating costs. Fixed-volume savings and additional completions are alternative views, not a combined promise. Replay performance is not independently verified task success; Infaio queueing and client-to-ATOM latency are not modeled.
Workforce-time value
Hours returned = baseline runs × workflow seconds saved ÷ 3,600 × recoverable-time share. Value = hours returned × hourly workforce cost. Defaults are $100/hour and 100% recoverability, representing one person blocked on each run who can use all the saved time.
Reduce recoverability for unattended agents, overlapping waits, or time that cannot be reused. This estimates productive capacity, not automatic payroll savings. It is calculated on baseline runs only and remains separate from inference savings.