The models can be very optimized for the harness, Clem (of Hugging Face) just recently released a Qwen3.8 model that used RL training on the harness and it got some pretty significant gains in the benchmarks.
- SWE-bench Pro (hard 51): 39.2% → 64.7%
- Terminal-Bench 2.1: 39.5% → 55.8%
I'll call bullshit on that.
If this were true, models would do way better in their own harness than in others, which doesn't seem to be the case (effect is minor).
Your link has a "link to the release blogpost" which is supposed to explain the training recipe but the link is down.