GPT-5.4

This is the strongest throughput case at 92.7 tok/s median with no issues, indicating good sustained generation on structured, mid-length outputs. The 3,410-token average output did not materially degrade speed.

Throughput is stable at 88.0 tok/s median, essentially matching the overall median and indicating consistent decode behavior on technical content. Output length is slightly shorter than API Design, but there is no sign of instability.

This is the main throughput weakness at 57.0 tok/s median, a large drop versus the other throughput prompts. The 3,571-token average output suggests longer-form prose is more expensive for this model and likely drives the lower overall average throughput.

TTFT is slightly slower here at 406 ms median, but still in a good range for interactive use. The very small 58-token outputs keep the latency profile focused on startup rather than decode.

This is the fastest startup case at 387 ms median TTFT, indicating low first-token latency on short factual responses. The 12-token average output is tiny, so this prompt is a clean read on initiation speed.