Operad cleared 183 of 183 levels across all 25 public ARC-AGI-3 games.
Operad, our internal agent harness, won 25 of 25 games with a mean RHAE of 99.60, against 96.78 for Codex (VISTA), and used 53% fewer actions than the human baseline.
Progress and efficiency
Levels and actions
Actions per game
Tokens
Input tokens per game
Output tokens per game
Input includes cached tokens; output includes reasoning tokens. Games ordered by Operad’s input tokens. Percentages are Operad relative to Codex (VISTA).
Games
| Frame | Game | Levels | RHAE | Actions |
|---|
Bars in each column share one scale; RHAE runs from 0 to 100. For actions, tokens and time, shorter is better. Select a game for its levels and the agent’s own notes.
RHAE is ARC-AGI-3’s relative human action efficiency: per level, (human actions ÷ agent actions)², capped at 115%, weighted by level and capped at 100. One scored run per game, GPT-5.6 Sol at maximum reasoning, 2,000 actions per game, scored offline with the official formula. The Codex (VISTA) figures come from our replication run of the published VISTA harness on Codex with the same model; VISTA’s own published ARC Prize scorecard for this setup is 99. Input tokens include cached tokens; output tokens include reasoning tokens. Time runs from a game’s first to its last agent event. The progress chart aligns every game to its own start and sums levels and actions across the 25 games. Snapshot 2026-10-13.