Grok 4.20 Reasoning

This is the strongest throughput case at 258.7 tok/s median with no issues, indicating the model can sustain high decode speed on long-form technical generation. The 2666 average output tokens suggest it holds performance well over extended generations.

Throughput is still strong at 237.7 tok/s median, close to the overall median, with no failures or anomalies. Output length is shorter than API Design, but the model remains stable and consistent.

This is the weakest sustained-throughput prompt at 172.6 tok/s median, a notable drop versus the other throughput tasks. The slower rate on a technical essay suggests longer-form prose generation is more decode-limited for this model.

TTFT is relatively slow at 1551 ms median, which is worse than the factual TTFT case and close to the overall TTFT average. The very short 71-token outputs mean startup latency dominates the user-visible delay here.

This is the best startup case with 932 ms median TTFT, indicating the model can begin responding quickly on short factual prompts. The 13-token average output keeps decode cost minimal, so this prompt is mostly a pure latency test.