China’s DeepSeek beefs up agentic AI with ‘harness’ tests as V4 model jolts Silicon Valley

Chinese AI firm DeepSeek is intensifying its push into agentic artificial intelligence (AI) technology by inviting open-source developers to test its upcoming "harness" software. Agentic frameworks, designed to help large language models (LLMs) autonomously execute multi-step coding tasks and complex workflows, have become a key area of competition among AI labs following the success of Anthropic's Claude Code.

DeepSeek, based in Hangzhou, is currently seeking beta testers for its DeepSeek Harness framework. The harness was introduced by Cui Tianyi, who leads the harness team and joined the company in March after leaving Hong Kong-based TSY Capital. The team is actively recruiting, despite being understaffed, according to Cui, who acknowledged the ambitious goals but noted the shortage of personnel.

Meanwhile, DeepSeek's latest V4 Flash model, "V4 Flash 0731," was released on Friday. This lightweight version, with 284 billion parameters, offers near-frontier performance at a significantly lower cost compared to competitors. The model has outperformed the OpenAI GPT-5.6 Luna in benchmark tests, matching its performance while being 60% cheaper in application programming interface (API) pricing.

On average, the DeepSeek model costs just US$0.03 per task, making it approximately 105 times cheaper than Anthropic's Claude Fable 5, which is priced at US$3.15 per task.

This cost-efficiency advantage has sparked concerns in Washington and Silicon Valley, as Chinese open-weight models gain widespread adoption. DeepSeek's strategy focuses on providing highly capable models at a fraction of the cost, aiming to achieve artificial general intelligence (AGI) – a goal where AI systems can match or surpass human cognitive abilities.