DeepSeek has moved its V4 Pro model out of testing while also releasing its agent software, Harness v0.1, under the MIT license. The Decoder reports that the new deepseek-v4-pro endpoint keeps the same model name, parameter count, and one-million-token context window, so existing integrations should continue working.
The update adds native support for the OpenAI Responses API with Codex integration and lets users set reasoning effort to low, high, or max. DeepSeek recommends the middle setting for everyday agent use. In its own comparison table, the company reports large gains on Terminal Bench 2.1 and DeepSWE.
The release also changes the economics. DeepSeek is raising API prices, with cache hits increasing to six times their current cost. That matters for agent workflows that repeatedly read the same files, because cached context can become a major part of total spending.
Independent analysis cited by The Decoder still places V4 Pro behind several top models on an intelligence index. The news is therefore not simply that DeepSeek has the strongest model, but that it is packaging a stronger model, agent tooling, and new pricing at the same time.