VISTA: Claude Opus 5.0 clears 25 ARC-AGI-3 games faster than people

A team including Kaiming He posted a paper on arXiv on 1 October 2026 describing VISTA, a harness that gives a multimodal model lossless visual memory. With it, Claude Opus 5.0 completed all 25 public ARC-AGI-3 games using 57.4% fewer actions than first-time human players.

Key takeaways

Paper posted to arXiv on 1 October 2026: VISTA, a visual harness for interactive reasoning
Claude Opus 5.0's Relative Human Action Efficiency rises from 40.68 to 100.00
The model finished all 25 public ARC-AGI-3 games
It used 57.4% fewer actions than people playing for the first time
Authors: Qiushi Han, Keya Hu, Linlu Qiu, Cathy Wu and Kaiming He

40.68 → 100.00Relative Human Action Efficiency for Claude Opus 5.0 once the VISTA harness is attached — with no change to the weightsarXiv 2610.02200

Memory that does not summarise

A typical harness describes observations in words and keeps a condensed transcript. VISTA does the opposite. The model sees the environment directly, and every earlier observation stays in its original form rather than compressed into text. It can reach back for any of them and reorganise its visual input while reasoning. The authors call this lossless visual memory.

VISTA

Observation from the environment

Stored in lossless visual memory

Does the model need an earlier observation?

YES

Retrieves it and rearranges its inputAllow

Works from the current view

Action in the environment

What ARC-AGI-3 measures

ARC-AGI-3 is an interactive benchmark, not a set of single-answer tasks. The agent lands in an unfamiliar environment, works out the goal itself and builds a world model as it plays. Scoring asks about efficiency, not victory alone: according to ARC Prize, 100% means beating every game as efficiently as a human.

\dots ▶ Symbol meaning \dots actions spent by the model running the VISTA harness \dots actions spent by people playing for the first time \dots the share of a human's actions the model needs for the same result

The jump from 40.68 to 100.00 therefore signals a move from trial-and-error to planned play rather than a sudden grasp of the rules.

Same model, different wrapper

The central claim concerns where credit sits. The underlying model is unchanged and only the way observations reach it differs. Across three further benchmarks of visual games and puzzles, VISTA clearly beats variants of the same model running minimal harnesses. That suggests some of today's agent failures come from the wrapper rather than from missing capability.

Why it matters

The industry usually answers weak agent results with more training. This work points to a cheaper route: leave the weights alone and fix how the model receives and retrieves observations. If the result holds, interactive benchmarks will start measuring the model-plus-harness pair rather than the model. That also changes comparisons, because a number without a harness description stops being comparable. For teams building agents, auditing the harness becomes as important as picking the model.

What's next?

The paper is a preprintPreprint: A research paper released publicly before peer review. Results can still change or fail verification. and has not been peer reviewed, so the 100.00 figure needs independent replication on the public game set
The public ARC-AGI-3 set covers 25 games and the private set stays out of reach, which limits conclusions until a hidden-set run happens