BlenderBench · hazmat retreat escape
Recorded agent trajectories.
20 Blender tasks · 40 saved model attempts · GPT-6 Astra and Claude Opus 5.5. Each task groups both models. Accepted verifier-only regrades preserve original traces and submitted scenes. Scores are continuous rewards, not binary success rates. This release is for benchmark evaluation.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.
Recorded agent trajectories.