DeepSeek quietly released the formal version of its V4 Pro model on August 13, with the model version updated to DeepSeek-V4-Pro-0813 across both API and Chat interfaces. The model name remains unchanged, allowing developers to call the latest version using "deepseek-v4-pro."
The V4 Pro formal version supports 1 million token context length and a maximum output of 384,000 tokens, making it particularly powerful for long codebase analysis, large-scale document processing, and multi-step agent tasks. It supports both thinking and non-thinking modes, with thinking mode enabled by default.
Benchmark results show dramatic improvements over the preview version. The DeepSWE software engineering benchmark score jumped from 12.8 to 62.7, while the Cybergym security agent benchmark rose from 52.7 to 83.3. The model surpasses Claude Opus 4.8 in multiple agent-related evaluations and rivals Claude Fable 5 in several categories.
The formal version also introduces native image reasoning capability for the first time, enabling the DeepThink engine to analyze images, interpret screenshots, and process mixed documents within a single reasoning flow.
Pricing for V4 Pro is set at 0.025 yuan per million tokens for cache-hit input, 3 yuan for cache-miss input, and 6 yuan for output - approximately three times the rates of V4 Flash. DeepSeek had previously announced plans to significantly increase API pricing across the board, though specific adjustments have not yet been implemented.
Source: TechWeb / Beijing Daily, August 13, 2026
