Deepseek has released an updated version of its V4-Pro model, which shows improvements in agent benchmarks but still lags behind top models like Claude Opus 5. The company also announced higher API prices and open-sourced its agent software. The new model, named Deepseek-v4-pro, is available under 'Expert Mode' and features native support for the OpenAI Responses API with Codex integration. Users can adjust reasoning effort to three levels: 'low,' 'high,' and 'max,' with the middle setting recommended for everyday use. According to Deepseek's comparison table, Terminal Bench 2.1 scores increased from 72.1 to 87.9, and DeepSWE scores rose from 12.8 to 62.7. On several agent benchmarks, the model outperformed Claude Opus 4.8. The update also addresses the growing competitiveness of the smaller V4 Flash model, which has narrowed the gap with the flagship product. Deepseek Harness v0.1, an open-source agent software, is now available as a Developer Preview under the MIT license. It is built on the Cordis plugin system, allowing for modular features like tools and sandboxes. The software includes a continuous session log and supports minimal mode for benchmarking. The project is led by Cui Tianyi, who joined Deepseek from Jane Street in March 2026. The company also announced new API pricing, effective August 16, with peak and off-peak rates. Off-peak usage costs half as much as peak hours, which run from 1 a.m. to 4 a.m. and 6 a.m. to 10 a.m. UTC. For V4-Pro input, off-peak rates increase from $0.435 to $0.66 per million tokens, and output jumps from $0.87 to $1.98. Cache hits, which are now the most expensive part of the change, will cost $0.022 off-peak and $0.044 at peak. The price hike partially reverses a previous reduction in May. The changes come as Deepseek raises new capital and prepares for an initial public offering.
Source: thedecoder
