#Terminal-Bench

Same Model, 22-Point Score Gap: Agent Harness Is the Key Variable

Same model, different harness: Terminal-Bench 2.0 scores differ by 22 points—Opus 4.6 from 60.45% to 82.02%, GPT-5.5 from 66.07% to 88.09%. A look at agent harness architecture, SDKs, and a minimal Python loop.