DeepSeek pushes AI programming into the 'prompt era': Creates 3D games from scratch, DeepSWE benchmark outperforms Opus 5

Dynamic Insight’s Beating AI News Flash: DeepSeek has upgraded its V4.1 Flash model to version 1017, with updates rolled out simultaneously for its web platform, app, and API. The new model weights have been open-sourced under the MIT license. Official test results show that DeepSWE v1.1 software engineering task pass rate rose from 74.2% to 77.0%, surpassing Claude Opus 5’s 74.0%. Terminal-Bench 4.0 improved from 31.2% to 42.3%, though still lagging behind Opus 5’s 51.8%. Per DeepSeek’s published cost curve, the configuration achieving the 77% pass rate costs an average of approximately 1.3 yuan per task. It’s worth noting that the comparison benchmark here is Opus 5, not the newer Opus 5.5. A key focus of the new version is enhancing "wish-based programming" – a feature where users only provide simple requirements, and the AI independently plans and executes the entire project. DeepSeek improved this capability via reinforcement learning with self-feedback, showcasing demos including a 3D Forbidden City scene, a web-based racing game, and a Godot engine-based recreation of Portal 2. The company stated the new model reduces both task completion time and output length, with optimizations tailored for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Additionally, DeepSeek is opening its self-developed search engine’s Search and Fetch APIs to developers for the first time. Developers can now enable AI agents to directly search web pages and extract content. Both APIs are billed per successful request, at 5 yuan per 1,000 requests. Search requests triggered via Claude Code and Codex’s built-in web search function will be free of API call fees for the first two weeks.