Overview
- DeepSeek has published model Deepseek-V4-Pro-0813 on its API and pricing pages, making the Pro build available for API calls under that name.
- The model interface advertises a 1,000,000‑token context window, up to 384,000‑token outputs, explicit tool-call support, structured JSON output, and compatibility with OpenAI and Anthropic API formats.
- Reporting says V4 Pro uses a mixture‑of‑experts (MoE) architecture with vendor‑reported totals of about 1.6 trillion parameters and roughly 49 billion active parameters per inference, but those architectural figures have not been independently verified.
- DeepSeek published per‑million‑token rates for Pro (cache‑hit input 0.025 CNY, cache‑miss input 3 CNY, output 6 CNY) and set a 500 concurrent‑request cap per account, and it also warned on the pricing page that API prices will be raised in the near term.
- Media tests cited by coverage show large benchmark gains over the V4 Pro preview on long‑task, code and security evaluations, and the launch follows the July release of the cost‑optimized V4‑Flash as DeepSeek shifts Pro toward complex, lower‑frequency, high‑value agent use cases.