DeepSeek has paired a broad product rollout with its previously flagged API price increases, releasing DeepSeek-V4-Pro, publishing DeepSeek-V4-Pro-0813 on Hugging Face with downloadable weights under the MIT license, and launching DeepSeek Harness as its first agent product. The company also provided vLLM, SGLang and local deployment instructions, allowing developers to self-host the new version. The update confirms previously leaked evaluation results for the DeepSeek-V4-Pro Agent and pushes DeepSeek further into the agent-runtime race. DeepSeek Harness, now available on npm, supports project management, long-running tasks, multi-agent workflows, skills, web search and context management, while letting users create different agent presets for coding, research and writing with different tools, prompts and operating modes. DeepSeek said Harness is built around an "Everything is a plugin" architecture, making models, tools, prompts, storage, context and even the interface replaceable or composable through plugins. Its underlying Cordis framework was released alongside a design paper that breaks model adapters, the tool system, Session Log and the Agent Loop into interchangeable components that can be loaded, removed and combined at runtime. The launch quickly drew developer attention. Roughly 90 minutes after DeepSeek Harness was open-sourced, its GitHub repository had surpassed 22,000 stars, a pace that the report said exceeded the previously public records set by xAI Grok-1 and DeepSeek-R1 in reaching 20,000 stars. The release also lands as DeepSeek adjusts V4 pricing, including a 4.5-fold increase in DeepSeek-V4-Pro peak output pricing from earlier levels, with some analysts saying the company's historical low-price advantage is being repriced and that peak Flash pricing is now above GPT-5.6 Luna.