DeepSeek launched DeepSeek Harness, an open-source tool that converts language models into agents capable of working directly on software projects, positioning the company to compete more directly with solutions like Claude Code. Available initially as a developer preview under an MIT license, the project embraces a "Everything is a plugin" philosophy—meaning models, tools, skills, sessions, isolated environments, file systems, and even execution cycles can be replaced or extended through plugins. Built on Cordis, Harness lets developers create more customized agents rather than relying on closed frameworks tied to a single model. DeepSeek notes the project evolves rapidly and may receive breaking updates.
The launch coincides with the official arrival of DeepSeek V4 Pro on its API and a significant pricing shift. Starting August 16, pricing varies by usage time. V4 Pro's current rate of $0.435 per million input tokens (non-cached) and $0.87 for output moves to $0.66 and $1.98 off-peak, then climbs to $1.32 and $3.96 during peak hours—meaning even the cheapest tier now exceeds previous prices. This V4 Pro and Harness combination signals a strategic pivot: rather than compete solely on cheap models, DeepSeek is building a complete platform for agent-driven development, entering a crowded market where Anthropic, OpenAI, and others increasingly vie for developer workflows.