i am being a bit pedantic, sure, but part of my point is that codex and claude code are absolutely general, the fact that they aren't behind the API layer is irrelevant. (it's path-dependent compatibility because providers hew to the completions API. claude -p is a general mechanism.)

if they plan to redo the evals on these harnesses (or if I missed them doing that already) that's one thing. but i saw someone had vanilla claude cowork solving the puzzles pretty well, so...