DeepSeek V4 Flash 0731: Reasoning skills rival big players, but costs only 4 cents DeepSeek V4 Flash 0731 achieved 61.4% accuracy in abstract reasoning tests (ARC-AGI-2), with a cost of just 4 cents per task—over 10 times cheaper than closed-source models like OpenAI and Anthropic. This confirms a key insight: open-source models, through the 'compute-at-inference' approach, are closing in on big players' performance at a much lower cost.