- SWE-bench Pro (hard 51): 39.2% → 64.7%
- Terminal-Bench 2.1: 39.5% → 55.8%
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.
For what it’s worth, the linked article data (Agent Duel data) is all run on OpenCode. I’m curious if people would prefer we run each model in its native harness (at the expense of user experience because it’ll be buggy).
My company is slow-walking (and that is an understatement) our LLM usage, so the majority of work that I'm reviewing is human generated. After the stories I hear of what other people are experiencing when management is more enthusiastic, maybe I should count myself lucky.
"OK, google models perform just as well, are faster and are cheaper but all that doesn't matter. Because the models, now that Google has very good ones, doesn't matter after all. All that matters is the harness".
And of course it's upvoted by the Anthropic shills.