MIT's VISTA harness lifts Claude Opus 5.0 to a perfect 100 RHAE on ARC-AGI-3
MIT CSAIL's Kaiming He team published VISTA on October 1 — a visual harness with lossless visual memory that raised Claude Opus 5.0's Relative Human Action Efficiency on ARC-AGI-3 from 40.68 to 100.00. The model cleared all 25 public games using 57.4% fewer actions than first-time human players, with no training and a four-sentence prompt.
- RHAE rose from 40.68 to 100.00, clearing all 25 public games
- Claude used 7,302 steps — about 0.43x the human average
- A 320B open-weight model jumped from 1.89 to 66.93 via the harness alone
- Extending context from 200K to 780K dropped the score from 99 to 93.9
Read next
AI