DeepSeek V4 Flash 0731

DeepSeek V4 Flash 0731 achieved 89.0% accuracy on ARC-AGI-1 and 61.4% on ARC-AGI-2 benchmarks at low costs of $0.02 and $0.04 per task respectively when operating at maximum effort with three reasoning variants. The model demonstrates varying performance across different reasoning levels (Max, High, Low) and task complexities, with detailed results provided across 400 tasks in ARC-AGI-1 and 120 tasks in ARC-AGI-2.