DeepSeek Releases V4-Pro and Harness: Flash Peak Price Surpasses GPT-5.6 Luna, Harness Sets New GitHub Star Growth Record

BlockBeats news, on August 14, DeepSeek's product line underwent a series of intensive updates, officially launching DeepSeek-V4-Pro and introducing its first Agent product, DeepSeek Harness. At the same time, DeepSeek-V4-Pro-0813 was officially uploaded to Hugging Face, with weights available for download under the MIT license. The official release also includes vLLM, SGLang, and local deployment instructions, enabling developers to deploy this version independently. In terms of product capabilities, previously leaked evaluation results for the DeepSeek-V4-Pro Agent have been confirmed. The official version of DeepSeek-V4 also saw a price adjustment, with the peak output price for the Pro version increasing by 4.5 times compared to earlier rates. Analysts suggest that after the price hike, DeepSeek's Flash peak pricing has surpassed GPT-5.6 Luna, indicating that its low-cost advantage is being redefined. The Agent direction is the core focus of this update. DeepSeek Harness has been released on npm as DeepSeek's first Agent product, supporting project management, long tasks, multi-Agent operations, skills, online search, and context management. It also allows users to create different Agent presets for tasks such as coding, research, and writing, with configurable tools, prompts, and workflows. Architecturally, DeepSeek Harness adopts "Everything is a plugin" as its core design principle, enabling models, tools, prompts, storage, context, and even interfaces to be replaced or combined via plugins. Its underlying framework, Cordis, has also been made publicly available through a design paper, breaking down model adapters, tool systems, session logs, and Agent loops into interchangeable components that support runtime loading, unloading, and combination. Market enthusiasm is rapidly heating up. Within approximately 1.5 hours of DeepSeek Harness being open-sourced, its GitHub Stars surpassed 22,000, breaking the previous record set by xAI Grok-1 and DeepSeek-R1 for reaching 20,000 Stars. Industry attention has shifted from focusing solely on individual model capabilities to exploring the engineering foundations required for Agent runtime, plugin-based systems, and recursive self-improvement.