DeepSeek V4 Flash 0731 — ARC-AGI Results
DeepSeek V4 Flash 0731 is a reasoning model with three effort variants (Max, High, Low), submitted to the ARC Prize leaderboard. At max effort, it scores 89.0% on ARC-AGI-1 (Semi-Private) at $0.02 per task and 61.4% on ARC-AGI-2 (Semi-Private) at $0.04 per task. No ARC-AGI-3 scores are reported.
Verified Scores
| Variant | ARC-AGI-1 | ARC-AGI-2 | ARC-AGI-3 |
|---|---|---|---|
| Max | 89.0% | 61.4% | — |
| High | 87.0% | 56.0% | — |
| Low | 84.0% | 46.0% | — |
The gap between effort levels is modest on ARC-AGI-1 (5pp from Low to Max) but substantial on ARC-AGI-2 (15.4pp), suggesting the harder benchmark is where reasoning depth matters most. The cost is extremely low by frontier standards — $0.02–0.04 per task makes this one of the most cost-efficient high-scoring entries on the leaderboard.
The public eval tables show per-task pass/fail across all three variants for 400 ARC-AGI-1 tasks and 120 ARC-AGI-2 tasks. The pattern is consistent: Max passes more tasks than High, which passes more than Low, with no task ever passing at a lower effort level but failing at a higher one (the reasoning-effort monotonicity holds).