(Yicai) Oct. 19 -- Chinese artificial intelligence startup DeepSeek has rolled out an upgraded AI model that runs faster and costs less to use, while opening its web search tools to outside developers to integrate into their own applications.

The Hangzhou-based company recently released V4.1-Flash-1017, following the debut of V4.1 Flash on Sept. 10. The update improves reasoning, coding, agent tasks, and visual performance, according to the firm, known for its low-cost options to ChatGPT and Claude.

DeepSeek described the release as a “minor update” in its technical report. The new large language model also generates shorter responses, helping users complete tasks faster and at lower cost.

DeepSeek also made its proprietary Search and Fetch interfaces, which allow developers to search the web and retrieve content from webpages, publicly available through its application programming interface. Both services cost CNY5 (75 US cents) per 1,000 calls.

The company said its search engine is designed to deliver relevant, authoritative, and timely results across multiple languages while filtering out search engine optimization spam, generative engine optimization (GEO) manipulation, and low-quality AI-generated content, aiming to provide reliable, up-to-date results.

According to DeepSeek's benchmark results, V4.1-Flash-1017 topped several tests, including the AI coding benchmark DeepSWE v1.1, the agent task completion benchmarks AutomationBench and Agents' Last Exam, and the multimodal agent benchmark Apex Bench v1.1. The model outperformed overseas competitors Claude Opus 5 and GPT-5.6 Sol, as well as Chinese rivals Kimi K3 and Zhipu's GLM-5.3, in those tests, according to the firm.

DeepSeek also updated DeepSeek Harness, a tool that helps AI models carry out complex tasks, to version 0.2.1 to work with the latest release. The new model was optimized for Harness during training to improve task performance while reducing runtime, output length, and token costs.

Editor: Emmi Laine