DeepSeek-V4-Pro Officially Released: Previously Leaked Agent's Skyrocketing Performance Fully Validated

According to Sentinel Beating monitoring, DeepSeek has officially released the V4-Pro final version, with the app, web interface, and API being updated simultaneously. The previously leaked Agent evaluation scores have also been officially confirmed: Terminal Bench 2.1 scored 87.9, DeepSWE jumped from the Preview score of 12.8 to 62.7, and CyberGym scored 83.3.

The new version also natively supports the Responses API and has been adapted for Codex. The thinking intensity is now divided into three levels: low, high, and max, which can be adjusted according to the task difficulty.

The price remains unchanged for now, but DeepSeek has announced that starting from August 17, the API pricing will be adjusted to adopt a peak/off-peak pricing model. The off-peak price will be only half of the peak price, and the specific new prices have not been disclosed yet.