On the evening of August 13, DeepSeek officially announced the launch of the DeepSeek V4 Pro full version, a move that had been circulating in the industry the previous night. More notably, the announcement detailed a significant adjustment to API pricing, a change the company has been planning for some time, with earlier warnings on its website indicating a "substantial increase."
Developers report that the pricing changes are drastic. The overall adjustment for DeepSeek V4 Pro results in prices ranging from 1.5 times to 12 times the original levels, translating to an increase of 0.5 times to 11 times. DeepSeek stated that the API will adopt peak-valley pricing, with off-peak rates set at half the peak rates. Peak hours are defined as 9:00-12:00 and 14:00-18:00 Beijing time, with all other times considered off-peak. The new pricing takes effect on August 17, 2026, at 00:00 Beijing time.
To illustrate the changes, for the newly released DeepSeek V4 Pro, the current price per million tokens is 0.025 yuan for input (cache hit) and 3 yuan (cache miss), with output at 6 yuan. After August 17, peak-hour prices will be: input at 0.3 yuan (cache hit) and 9 yuan (cache miss), with output at 27 yuan. Off-peak prices will be: input at 0.15 yuan and 4.5 yuan, with output at 13.5 yuan. Calculated at the highest peak rate, the input price per million tokens for DeepSeek V4 Pro will be 12 times the original (cache hit) and 3 times (cache miss), representing increases of 11 times and 2 times, respectively, while the output price becomes 4.5 times the original. Even during off-peak hours, the three price categories will be 6 times, 1.5 times, and 2.25 times the original.
DeepSeek stated that this peak-valley pricing is designed to "allocate resources more reasonably" and encourages users to adjust task timing based on actual usage. The price increase does not affect free users of the new model. Both web and app users can now select "Expert Mode" to experience the full version of DeepSeek V4 Pro. The real cost recalculation impacts developers and enterprise clients who integrate the model into their products via the API.
"It's no longer cheap; the price is now on par with similar models," one developer commented, noting that the impact is significant, especially with the sharp peak-hour increases. "This Pro version lacks cost-effectiveness." Another user called for a rapid influx of new Shenghui chips to drive down DeepSeek's prices again.
Regarding the updated DeepSeek V4 Pro, the company stated that its agent capabilities have been enhanced, leading to noticeably improved performance in production environments. The V4-Pro and V4-Flash thinking modes now support three levels of thinking intensity: low, high, and max. Users can select based on task complexity: low for simple tasks, high for everyday agent tasks, and max for highly complex scenarios.
Preliminary evaluations from overseas AI model testing agencies indicate that the new model ranks among the top tier of mainstream models but does not lead across the board. According to the AI model testing agency Artificial Analysis, the latest DeepSeek V4 Pro scored 53 in intelligence, ranking 9th on the list. Based on Arena.ai scores, the current DeepSeek-V4-Pro (Max) ranks approximately 8th overall, and among open-source models, it is second only to Kimi K3 (Max) (with a score of 1674 points). However, these rankings are based on limited user voting and may change.
Historically, low pricing was a key advantage for DeepSeek. This price increase signals that it may no longer be affordable, but whether it offers value for money ultimately depends on performance. The crucial question is whether the new model can enable developers to accomplish more work with fewer tokens in real production environments, determining if they can accept the higher bills.