DeepSeek V4 Pro has been online for two days, and I have reorganized all the current controversies. That's all for now After the release of DeepSeek V4 Pro 0813, there have been almost two models arguing online. On one hand, there is the official exaggerated Agent score, while on the other hand, a large number of developers claim that it is not at all based on this physical sensation. 1. The publishing process is very chaotic On the evening of August 12th, the API documentation and model version underwent changes first. DeepSeek officially announced the V4 Pro GA on August 13th. There have been instances where banners and announcements on the official website have been withdrawn or re launched, but DeepSeek has yet to explain what exactly happened during this process. More sensitive is that the price has also changed accordingly. The new price will take effect on August 17th, introducing peak valley pricing, and the peak output price of V4 Pro will reach $3.96 per million tokens. That is to say, while users are still judging how much stronger the official version is, they are already facing significant price increases. 2. The official score does not match the user experience of a large number of users This is currently the most controversial area. The official data for the V4 Pro's Agent is very impressive: tasks such as DeepSWE 62.7, CyberGym 83.3, and Terminal Bench have also reached the top level. But the comprehensive Intelligence Index given by third-party Artificial Analysis to 0813 Max is 53, and there has been no equivalent level of crushing upgrade. The most intense feedback from the community in the past two days has focused on a few issues: Short thinking in complex tasks, premature termination of agents, and unstable execution of long tasks. Sometimes it even feels inferior to Flash, and the difference in results after changing the time and session of the same Prompt is significant. 3. V4 Pro may heavily rely on Harness/Scaffold This is the most noteworthy aspect in my opinion. DeepSeek has actually publicly admitted to using DeepSeek Harness's Minimum mode in some Code Agent benchmarks. This mode is not simply changing the Prompt, but rather integrating the System Prompt, number of tools Tool Schema、 Context compression methods, etc. should be made as familiar as possible in the Agent environment during the model reinforcement learning stage. The community subsequently conducted some interesting experiments on V4 Pro 0813. For the same task, DSH Standard/PTC scores approximately 91-92 points, but when switched to Minimum, it can reach 96-99 points. What's even more exaggerated is Anchored Standard. In the first round, we used Minimal to bring the model in, and after the first Tool Call, we restored the complete tool. We were able to run 98-99 twice. This set of tests is not yet large enough to prove that all tasks are like this, but it at least exposes one issue: The ability of the V4 Pro may not be a fixed 90 or 60 points, but rather very impressive. 4. There are multiple models behind the hottest API Some people in the community suspect that the deepseek-v4-pro API actually routes multiple checkpoints behind it. The reason is that some people have found that after changing IP addresses and creating new sessions, the model's way of thinking will show different fingerprints: Some crazily write 'Let me...', some are used to 'The user wants me...', and there is another type that uses' we 'extensively and has significantly stronger abilities, so there are people who call them Preview version, Flash version, and God version. This claim has spread very quickly, but there is currently no official evidence to prove that DeepSeek secretly switches multiple checkpoints after a model name. What may be more explanatory now is the combination of two layers: differences in server grayscale/cluster configuration, and extreme sensitivity of V4 Pro to the first round of System Prompt, Tool Schema, and Agent Scaffold. So the effect that users see is like 'randomly drawing models'.