Gemini 3.1 Pro

This is the fastest throughput case at 127.1 tok/s median, indicating the model can sustain high decode speed on long-form technical output. With 2563 average output tokens and no issues, performance here is stable and representative of the model's best sustained rate.

Throughput stays strong at 123.6 tok/s median, only slightly below the API design prompt, so decode speed is consistent across technical prompts. The 2252-token average output length did not introduce any reliability or slowdown issues.

This prompt is the main throughput weakness, dropping to 100.4 tok/s median, about 20% below the other throughput tasks. The longer-form essay style likely increases generation complexity or variance, but it still completed successfully with no failures.

TTFT is slow at 7691 ms median even though the output is tiny at 63 tokens, so startup overhead dominates this case. This suggests the model's first-token latency is not well optimized for short responses.

This is the best startup case at 5739 ms median TTFT, but it is still materially high for a short factual response. The 31-token average output confirms the latency is mostly prefill/startup cost rather than decode time.