DeepSeek's new Flash model has matched OpenAI GPT-5.6 Luna at roughly 60% lower cost.

DeepSeek has released a new budget model, V4 Flash "0731", which scores 50 points on the Artificial Analysis Intelligence index, ten points more than the previous version released in April 2026. This is just one point less than OpenAI's budget model GPT-5.6 Luna, while the cost per task is about 60 percent lower, even accounting for OpenAI's 80 percent price cut. The key factor is DeepSeek's 98 percent cache discount, significantly higher than the industry average of 90 percent. The model also consumes twelve percent fewer tokens than its predecessor. Improvements span all test categories, especially agentic tasks: in the GDPval benchmark, which measures complex real-world office work, the score rose from 1189 to 1559 Elo points, and the hallucination rate decreased. The architecture remains unchanged: 284 billion parameters, of which 13 billion are active, with a context window of one million tokens. The model's weights are distributed under the MIT license on Hugging Face.