$5K LIBERO-10 baselines for GPT Astra, Claude Opus and Gemini-4

Yu Xiang @YuXiang_IRVLOriginally posted 2026.10.22

Open original on X

We spent $5K evaluating GPT Astra/Sol/Luna, Claude Opus, and Gemini-4 on LIBERO-10, so you can reuse our baselines for future experiments! All runs are zero-shot, with LLMs predicting end-effector poses. Testing how well LLMs can reason and act. Check May's thread below! 👇

This post quotes another post: “Six frontier models on LIBERO-10: success rate vs API cost”