ARC Prize reports DeepSeek V4.1 Flash 1017 at 78.1% on ARC-AGI-2

ARC Prize's results page reports DeepSeek V4.1 Flash 1017 scores on ARC-AGI benchmarks. Best Semi-Private scores are 94.5% on ARC-AGI-1 at high reasoning and 78.1% on ARC-AGI-2 at max reasoning. The page lists a verified scores table by reasoning level (max, high, low, none) and per-task pass/fail results on the ARC-AGI-2 Public Eval set of 120 tasks, alongside a leaderboard of many other models.

Key takeaways

• Best Semi-Private ARC-AGI-2 score is 78.1% at max reasoning.
• Best Semi-Private ARC-AGI-1 score is 94.5% at high reasoning.
• Scores are reported per reasoning level: max, high, low, and none.
• Per-task pass/fail results are shown across 120 ARC-AGI-2 Public Eval tasks.

Why it matters

Independent verification of frontier-model scores on ARC-AGI-2 helps readers compare reasoning-model progress on a benchmark designed to resist memorization. Per-level breakdowns show how much performance depends on reasoning budget.

Read with care

The supplied text is an excerpt of a results table; ARC-AGI-3 scores are shown as dashes for this model and the full methodology is not included.