How Powerful Is the DeepSeek V4 Official Version? The 'AI Execution Line' Lives Up to Its Name, and Its Value for Money Takes Out Most Competitors

Just in time for the last day of July, DeepSeek finally unveiled the official V4 version. The initial release was only the V4 Flash-0731; the official Pro version will arrive in early August, debuting alongside DeepSeek's Harness tool. After the V4 Flash-0731 release, netizens who had been calling Liang Wenfeng "Little Liang" and "Liang Baikai" due to DeepSeek's delays immediately restored his "Saint Liang" title, because its performance exceeded expectations. Official test performance surpassed GLM-5.2, slightly below Opus 4.8, with some tests even matching the latter. So just how strong is V4 Flash-0731? Many netizens have already conducted their own tests online, and the results largely match the official claims. Given its 284B parameters, its performance is nothing short of stunning, outperforming many trillion-parameter (1000B) large models. If you still can't grasp V4 Flash-0731's performance level, some netizens used GPT to analyze and aggregate data from the AI performance leaderboard Artificial Analysis, creating a detailed comparison combined with API pricing that covers nearly all common large models from China and abroad. This chart is simple: the x-axis is price (lower is better), the y-axis is performance score (higher is better), and each large model is also differentiated by thinking level. If you don't want to compare them one by one, just look at V4 Flash-0731's position. Any large model located in its lower-right quadrant means either lower performance than V4 Flash-0731 or worse pricing. In simple terms, those domestic and foreign large models in the lower-right have already been slain by V4 Flash-0731. The ones fortunate enough to escape its kill line are mainly two types: one is models priced lower than it, primarily GPT-5.6 Luna (low) and Xiaomi's MiMo-V2.5. The former just cut its price by 80%; otherwise it too would have fallen within the kill line. Xiaomi's large model had already cut prices following the V4 preview release, and it's also multimodal, giving it certain advantages. However, its coding performance lags a step behind the current V4 Flash-0731. It's speculated that Xiaomi will soon release an upgraded version. The large models in the upper-right that escape V4 Flash-0731's kill line are the flagship/mainstay models of current AI giants both domestically and internationally, including Kimi K3, Fable 5, Opus 5, Opus 4.8, GPT-5.6 Sol, and others. Moreover, they need their thinking level set to Max or High just to use performance advantages to pull away from V4 Flash-0731 and leave the danger zone. Overall, this chart clearly displays V4 Flash-0731's capability range. Netizens' earlier positioning of DeepSeek as the kill line in the AI field is quite an interesting and accurate concept. Whether domestic or foreign, there's no reward for models that surpass DeepSeek's performance — but if you can't even match DeepSeek's performance, its overwhelming price advantage will make things very uncomfortable for you. Competing on price is tantamount to profit suicide, and not competing on price means users will be siphoned away. And all of the above is just the capability of V4 Flash-0731, a lightweight model. DeepSeek's ace in the hole, the official V4 Pro, still needs a few more days, along with its companion Agent software. Considering its parameter count is roughly 5 times that of V4 Flash-0731, the performance is highly promising — it might even raise the kill line standard to the Opus 5/GPT-5.6 Sol level. Finally, V4 Flash-0731's naming has returned to last year's R1 style of keeping the version number unchanged while adding a date suffix. This means that over time, the post-training of both V4 Flash and Pro will continue to strengthen, with upgrades every two to three months. Competitors will have to keep up.