DeepSeek V4-Pro official version launches with peak-valley pricing

On August 13, DeepSeek announced via its official WeChat account that the V4-Pro official version is now live across the App, web, and API. Users can select the new model through the "Expert Mode" option. V4-Pro official version supports 1M context length and up to 384K token output, offering both non-thinking and default-on thinking modes. According to DeepSeek's published benchmarks, V4-Pro surpasses Anthropic Claude Opus 4.8 on HLE, Terminal Bench, Cybergym, and DeepSWE agent capabilities, and approaches Claude Fable 5 released in June, even surpassing it on some leaderboards.

Alongside the official release, DeepSeek unveiled its API pricing overhaul, adopting peak-valley pricing where peak-hour rates are double off-peak rates. Peak hours are defined as 9:00-12:00 and 14:00-18:00 Beijing time on weekdays. V4-Pro peak pricing is 0.3 yuan per million cached-hit input tokens, 9 yuan per million uncached input tokens, and 27 yuan per million output tokens; off-peak prices are 0.15, 4.5, and 13.5 yuan respectively. V4-Flash peak pricing is 0.1 yuan cached-hit input, 3 yuan uncached input, and 9 yuan output; off-peak prices are 0.05, 1.5, and 4.5 yuan. The new pricing takes effect at 00:00 Beijing time on August 17, 2026.

DeepSeek stated the mechanism is designed to allocate compute more rationally and encourages users to adjust task timing accordingly. The price increase does not affect free consumer access to the new model; developers and enterprises integrating via API will need to recalculate costs. V4-Pro and V4-Flash thinking modes support three intensity tiers (low, high, max), and DeepSeek API now natively supports OpenAI Responses API format with Codex-specific adaptations.