DeepSeek Enhances AI Programming Capabilities, Outperforms Claude Opus 5 in Software Engineering Tasks

Dongcha Beating AI News: DeepSeek has upgraded V4.1 Flash to version 1017, with simultaneous updates rolled out on the web, app, and API. The new model weights have also been open-sourced under the MIT license. In official tests, DeepSWE v1.1’s software engineering task pass rate increased from 74.2% to 77.0%, surpassing Claude Opus 5’s 74.0%. Terminal-Bench 4.0 improved from 31.2% to 42.3%, but still lags behind Opus 5’s 51.8%. According to official cost curves, achieving a 77% pass rate costs approximately RMB 1.3 per task on average. However, the comparison is with Opus 5, not the newer Opus 5.5. The new version significantly enhances “wishful programming”—where users provide only simple requirements, and the AI autonomously plans and completes entire projects. DeepSeek improved this capability using self-feedback reinforcement learning, demonstrating examples such as a 3D model of the Forbidden City, a web-based racing game, and a recreation of Portal 2 built with the Godot engine. The official team notes that the new model reduces both task completion time and output length, and has been optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Additionally, DeepSeek is for the first time opening its proprietary search engine’s Search and Fetch APIs to developers. Developers can now enable agents to directly search web pages and extract article content. Both APIs are billed per successful request at RMB 5 per thousand calls. Search requests triggered via Claude Code or Codex’s built-in web search functionality will have their API usage fees waived for the first two weeks.