DeepSeek V4 Flash Goes Public, Closing the Gap with GPT-5.6 at a Fraction of the Cost

On July 31, 2026, DeepSeek released the public version of its V4 Flash model. The model scored 50 points on the Artificial Analysis intelligence benchmark, just one point behind OpenAI's GPT-5.6 Luna. Only a day later, two leading evaluation platforms had already published fresh results. The launch showed that a Chinese open-source model can now sit shoulder to shoulder with the best Western closed models on raw quality. The price story is even more striking. Even after OpenAI cut the price of GPT-5.6 Luna by 80 percent, V4 Flash still costs about 60 percent less per task. The main reason is a deeper cache hit discount. DeepSeek gives users a 98 percent discount when the same question is asked again, while most of the industry only offers 90 percent. For startups and small teams, that difference adds up fast at the end of every month. V4 Flash is built on a Mixture of Experts architecture. The model has around 284 billion total parameters, but only 13 billion are active for any single question. This design keeps running costs low while still allowing long context windows of one million tokens. The full version, V4 Pro, is expected to follow in early August with even stronger coding and agent skills. The deeper meaning of this release is that the AI race is no longer only about who has the smartest model. It is also about who can offer the best balance of price, speed and quality. With DeepSeek planning a one gigawatt data center in Inner Mongolia and Chinese models already winning on cost, the gap between the top labs is shrinking faster than most experts predicted at the start of the year.