After the major price increase, computing power is now sufficient. There's a new version of DeepSeek V4: fresh evidence has emerged.

On August 19, after officially raising API prices several times on the 17th, DeepSeek's reputation has inevitably declined, but for them, this price hike isn't about profitability—it's due to insufficient computing resources, requiring focus on new model development.

Previously, DeepSeek's official assistant also explained that they would concentrate R&D efforts on delivering stronger models, and this isn't just talk—in fact, relevant evidence has already emerged.

Netizens testing have found that DeepSeek V4 Pro now features a new chain-of-thought, with responses starting using an "I'm doing" pattern, which differs from previous chains like "we need/I'll/Let's." This new chain-of-thought more closely resembles what was seen during earlier gray-scale testing.

Additionally, a Bilibili user posted a full screen-recording video proving that V4 has a new chain-of-thought, also similar to the gray-scale test phase, and welcomed fact-checking, given the recent flood of back-and-forth claims about DeepSeek V4.

Currently, it's unclear what specific performance changes the new chain-of-thought brings, possibly because test samples are still too limited and require further observation.

From these signs, DeepSeek V4, especially the V4 Pro model, likely has multiple weight versions, each with different capabilities. This explains why, after the 0813 release, different users reported varying V4 Pro performance—some found it exceptionally strong, while others found it mediocre, even inferior to the Flash model.

Given the various issues with last week's V4 Pro-0813 release, DeepSeek needs to focus on delivering a more stable V4 Pro version, especially after the API price surge. In recent days, due to the significant cost increase, call volumes for V4 Flash and Pro in China have plummeted, with severe user attrition.

Even if DeepSeek isn't overly concerned about C-end users, continuing this way won't work—they need to roll out compelling new models or features to restore confidence in DeepSeek.