Which Codex setup is fastest?
The same prompt, measured across models, reasoning efforts, and normal or fast service. I rerun the test locally and publish the results here.
Effective output speed. Reported output tokens divided by the full Codex CLI wall time, including startup, transport, reasoning, and orchestration.
Codex speed table
Every row is one provider × model × reasoning effort × service mode combination, ranked by median effective output tokens per second.
No benchmark has been published yet.
The next local run can be exported here without exposing prompts, outputs, or account details.
| Rank | Provider | Model | Reasoning | Mode | Median effective tok/s | TPS range | Median wall | TPS samples |
|---|
What “TPS” means here
Effective output tok/s is reported output tokens ÷ end-to-end Codex CLI wall time. It includes process startup, queueing, transport, reasoning, generation, and exit overhead, so it is useful for comparing the experience of these settings—not as a claim about raw server-side decoding speed. Warmups are excluded and repeat medians matter more than isolated runs.