Mercury 3 test: diffusion model is fast, but not uniquely so

An independent test of Inception's Mercury 3 across 48,000 API calls found it passes 86% of coding tasks at a 0.83 s median, while gpt-oss-120b on Cerebras hits 84.9% at 0.34 s. In a SQL agent Mercury 3 finishes tasks in 2.06 s versus 3.86 s for Claude Haiku 5.5.

• Mercury 3: 86% on HumanEval+, MBPP+ and LiveCodeBench at 0.83 s
• gpt-oss-120b on Cerebras: 84.9% at 0.34 s, faster than Mercury
• Mercury 3 streams 850–900 tokens/s versus 2,400 on Cerebras
• In the SQL agent Mercury 3 uses 5.1 turns, Claude Haiku 5.5 uses 3.2