DeepSeek Launches DeepSeek-V4-Pro: From Preview to General Availability

DeepSeek has officially moved its flagship model, DeepSeek-V4-Pro, from preview to general availability. The first version of the model’s architecture debuted in late April 2026, but after a major overhaul, the AI assistant now features a three-tier reasoning system and is available on chat.deepseek.com.

The key difference between the commercial release and the April preview is the introduction of flexible thinking modes in the API, allowing users to balance response quality and budget:

• Low: disables deep planning for fast responses with minimal token usage;
• High: standard planning mode for writing scripts, parsing data, and debugging;
• Max (V4-Pro-Max): uses the model’s maximum compute for multi-level programming and olympiad-level math.

The model’s context window is 1 million tokens, and the maximum generation limit per request has been expanded to 384,000 tokens, enabling the AI to produce architectural app designs in a single click. The release version also includes native support for the OpenAI Responses API and Codex configuration.

Benchmark Results

In the post-upgrade report, the model in maximum mode (V4-Pro-Max) achieved the following test results:

• LiveCodeBench (code generation): 93.5% success rate;
• SWE-bench Verified (GitHub issues): 80.6% resolved tickets (on par with Gemini 3.1 Pro);
• GPQA Diamond (logic and expert knowledge): 90.1% accuracy.

API Pricing and Dynamic Rates

The base API pricing for DeepSeek-V4-Pro is $0.435 per million input tokens and $0.87 per million output tokens. Starting August 16, the company will introduce dynamic pricing: during nighttime and off-peak hours, API request costs will automatically drop by 50%.