The general availability of DeepSeek-V4-Pro has been rolled out across the APP, Web, and API platforms. This update significantly boosts agent performance in production environments while introducing native support for the OpenAI Responses API format.

  • Agent benchmarks show HLE at 60.0 with tools and Terminal Bench 2.1 at 87.9.
  • The API now natively supports the OpenAI Responses API, adapted specifically for Codex users.
  • Thinking modes for V4-Pro and V4-Flash now offer three effort levels: low, high, and max.
  • Pricing adjusts to a peak/off-peak model where off-peak rates are half of peak-hour prices.

The release allows users to flexibly control thinking effort based on task complexity and provides a one-click configuration script for Codex integration.