ARC-AGI-3 SOTA; ~5x the next-best model (Opus 4.8 at 1.5%).

Terra scores 0.8% on ARC-AGI-3 vs Sol's 7.78% — abstract reasoning is Sol-only strength.

Score updated after ARC Prize changed scoring normalization (median player baseline, 115% cap per level). Was 0.2% under old scoring.

Highest score on new interactive reasoning benchmark. Humans score 100%. ARC-AGI-3 uses turn-based game environments with no instructions.

All frontier models below 1% on ARC-AGI-3. Interactive reasoning requiring exploration and continuous learning. Humans solve 100%.

Second-highest on ARC-AGI-3. All frontier models score below 1% on this new interactive reasoning benchmark.

Luna scores 0.18% on ARC-AGI-3; abstract reasoning reserved for Sol tier.

Only frontier model to score exactly 0% on ARC-AGI 3 at launch. Exceeded action cutoff on every level. Humans score 100%.