Claude Sonnet 4.6

This is the strongest throughput case at 101.4 tok/s median with 3 successful runs and no issues. Output length is large at ~4095 tokens, so the model maintains high decode speed on long-form technical generation.

Throughput is slightly lower than API design at 94.3 tok/s median, but still solid and stable across 3 successful runs. The near-identical 4096-token outputs suggest consistent long-context decoding without degradation.

This is the main throughput weakness, with median speed dropping to 48.8 tok/s across 2 runs. The long technical essay workload appears to stress sustained generation more than the structured technical prompts, pulling down the overall average throughput.

Short-definition prompts have the best startup latency at 986 ms median TTFT, which is the fastest TTFT in the set. With only ~80 output tokens, this suggests the model can begin responding relatively quickly when the completion is short.

Factual prompts are slower to start at 1428 ms median TTFT, the worst TTFT in the set. The low output length (~27 tokens) means the latency is dominated by prefill/startup rather than decode, so this is a clear first-token responsiveness weakness.