the same model in a different harness can yield much different performance! we've seen this on a few different occasions now - we took gpt-5…

the same model in a different harness can yield much different performance! we've seen this on a few different occasions now - we took gpt-5.2-codex from 52.8% to 66.5% on Terminal-Bench 2.0 (Top 30 to Top 5 at the time of publishing) just by applying harness layer changes like prompts and middleware hooks. - by following the specific prompting guides offered by frontier labs, you help the harness "speak the same language" as what the model was post-trained on. by modifying prompts, yet again, we found a 20% improvement from gpt-5.3-codex on tau2-bench (10% improvement for opus-4.7) details/posts ⬇️ @badlogicgames there's a lot of alpha in harness engineering there's a reason cursor's harness performs better on both claude and gpt models pi with a homemade apply_patch tool is great with gpt-5.5 for example, even if it doesn't match the exact codex cli/app spec :)

Source: Harrison Chase (X) | 2026-05-03