Roybench telemetry

Model speed history

Aggregate tokens_per_second is the arithmetic mean of output_tokens divided by the arithmetic mean of duration_seconds; it is not the arithmetic mean of per-attempt output_tokens / duration_seconds ratios. Effective output throughput includes proxy, network, provider queue, reasoning, and generation time; it is not raw decoder speed. Three Codex subscription GPT routes are post-response bounded at the declared 4,096 billed-output-token envelope. Eleven Cursor routes use provider defaults because the local agent exposes no documented output cap. Fifteen Devin routes send a 512-token request hint; throughput uses provider-reported output, and Devin documents no total-billed-output cap. The other eleven routes enforce a 512-token request output cap.