According to the official announcement, DeepSeek has upgraded its V4.1 Flash to version 1017, with web, App, and API endpoints synchronized. The new model weights are now open-sourced under the MIT license. Official benchmarks show that DeepSWE v1.1's pass rate on software engineering tasks increased from 74.2% to 77.0%, surpassing Claude Opus 5’s 74.0%; Terminal-Bench 4.0 score rose from 31.2% to 42.3%, though still below Opus 5’s 51.8%.

Configurations achieving 77% pass rate have an average cost of approximately ¥1.3 per question; the official benchmark comparison is against Opus 5, not Opus 5.5. The new release emphasizes "wish-based programming," enabling models to autonomously plan and complete projects based on simple requirements. Demonstrations include a 3D reconstruction of the Forbidden Palace, a web-based racing game, and a Godot-powered recreation of Portal 2. DeepSeek also unveils its proprietary search engine APIs—Search and Fetch—for the first time, both billed per successful request at ¥5 per thousand; Search requests triggered via Claude Code or Codex’s built-in web search will be free of API call fees for the first two weeks.