DeepSWE Author Confuses TPS with Pro System: Benchmark Comparison Misleading

teortaxesTex · x · 2026-08-14

stalkermustang criticizes DeepSWE author for confusing TPS with Pro (parallel agent system), citing winkeyh: comparing 'up to' throughput, Gemini 3.7 Flash is 500 TPS and GPT-5.6 Sol Fast 200 TPS. Thus 'up to 750 TPS' is impressive but misleading when plotted against Artificial Analysis's 72h P50 numbers.

Original post →

More from Models

Models channel →