DeepSeek Pushes AI Programming into the "Wishing Era": Building 3D Games from Scratch, DeepSWE Benchmark Score Surpasses Opus 5

Beating AI News Flash: DeepSeek has upgraded V4.1 Flash to version 1017, with the web, app, and API all switched in sync, and 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 rose from 74.2% to 77.0%, surpassing Claude Opus 5's 74.0%. Terminal-Bench 4.0 rose from 31.2% to 42.3%, but remains below Opus 5's 51.8%. According to the officially published cost curve, the configuration achieving a 77% pass rate costs an average of only about 1.3 yuan per question. However, the comparison target is Opus 5, not the newer Opus 5.5.

The new version focuses on strengthening "wish-driven programming," that is, giving only simple requirements and letting the AI plan and complete the entire project on its own. DeepSeek used self-feedback reinforcement learning to improve this capability and demonstrated a 3D scene of the Forbidden City, a web-based racing game, and a Portal 2 remake based on the Godot engine. The official statement said the new model reduced both the time and output length needed to complete tasks, and it has also been optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.

In addition, DeepSeek for the first time opened its self-developed search engine's Search and Fetch APIs to developers. Developers can let Agents directly search the web and fetch page content. Both interfaces are billed by successful requests, at 5 yuan per thousand requests. Search requests triggered through the built-in web search features of Claude Code and Codex will temporarily have their API call fees waived for the first two weeks.