DeepSeek will officially release DeepSeek V4 in mid-July 2026, introducing a new peak-valley pricing system that doubles API costs during defined peak hours, the company said [1, 2, 3, 4, 5]. Peak pricing will apply daily from 9 a.m. to noon and 2 p.m. to 6 p.m. Beijing Time, covering major workday hours [1, 2, 3, 4, 5].
The new system applies to two versions of the model. For the high-performance deepseek-v4-pro, the normal rates are ¥0.025 per million tokens for input cache hits, ¥3.00 for input cache misses, and ¥6.00 for output tokens. These prices will double during peak hours to ¥0.05, ¥6.00, and ¥12.00 respectively [1, 4]. The lightweight deepseek-v4-flash model has normal rates of ¥0.02 for input cache hits, ¥1.00 for cache misses, and ¥2.00 for output tokens; doubled to ¥0.04, ¥2.00, and ¥4.00 during peaks [1, 4].
DeepSeek explained peak-valley pricing aims to allocate computing resources more efficiently and improve service stability [2, 3, 4, 5]. The company began notifying developers by email on June 29, giving 24 hours’ advance warning of the pricing changes [1, 3, 4].
Some Chinese developers have criticized the peak pricing as unfriendly since it covers most of the regular workday. But others have welcomed the update, noting that true value lies in the capabilities of V4. One longtime DeepSeek API user said, "真正關鍵的不是價格，而是V4正式版的能力升級幅度" (The key is not the price but the capability upgrade in the V4 official release) [4].
DeepSeek V4 also introduces inference acceleration technologies called DSpark and DeepSpec to improve speed and lower costs [5]. The launch follows an April 24 preview release of V4 to select users [5], with the full commercial release and peak pricing set for mid-July 2026 [1, 2, 3, 4, 5].
