Grok 4

This prompt is near the model's overall median at 61.4 tok/s with 2354 average output tokens, indicating stable sustained generation on long-form technical content. No issues were reported, so throughput is consistent under this workload.

This is the fastest throughput case at 65.7 tok/s, suggesting the model handles structured technical exposition efficiently. The 2316-token average output is similar to API Design, so the higher rate is likely workload-dependent rather than due to shorter generations.

This is the weakest throughput prompt at 44.6 tok/s, a clear drop versus the other throughput tasks. The lower 1924-token average output still does not explain the gap, so this looks like a content-dependent slowdown on essay-style generation.

TTFT is slower here at 4134 ms, which is the worst startup latency among the TTFT prompts. The very short 67-token output suggests the model's initial response latency is the main cost, not decode length.

This is the faster TTFT case at 3286 ms, but startup is still multi-second and not especially responsive. With only 20 average output tokens, the latency profile is dominated by prefill/startup rather than generation.