DeepSeek V4-Pro Agent Benchmarks Surge 389% on DeepSWE

DeepSeek V4-Pro confirms leaked agent scores with 62.7 on DeepSWE and 87.9 on Terminal Bench 2.1. Peak-valley API pricing starts August 17.

Woofun AI reports that DeepSeek has officially launched the full version of V4-Pro, updating its application, web interface, and Responses API. The official team confirmed previously leaked agent evaluation metrics: the model achieved 87.9 on Terminal Bench 2.1, 83.3 on CyberGym, and a significant increase to 62.7 on DeepSWE from the preview version's 12.8.

The update introduces full support for the Responses API and Codex optimization, alongside three-tier thinking intensity settings. While current pricing remains static, DeepSeek announced the implementation of a peak-and-valley pricing structure for the API starting August 17, with off-peak rates set at half of peak hour costs.