DeepSeek has updated its Flash model to V4.1-Flash-1017, widening the cost-effectiveness gap on agentic tasks. On DeepSeek's own benchmarks, the model is more accurate and cheaper on programming agent tasks.
1. DeepSWE 77.0%, ahead of Claude Opus 5. On the DeepSWE v1.1 programming benchmark, the 1017 release scores 77.0%, above Claude Opus 5's 74.0%; the low-effort setting still reaches 72% at about $0.10 per task.
2. The secret is shorter outputs. The predecessor scored 74.2% at about $0.50 per task; the new version costs under $0.20 while gaining nearly 3 points of accuracy. For agents running hundreds of tasks a day, that difference adds up fast.
3. Open weights, and a search API. The weights are open on Hugging Face and ModelScope, and the model is available via the website, app, and API. DeepSeek has also opened its own search engine's search API to developers.
One caveat: these are DeepSeek's own benchmark numbers, and independent tests are not out yet. It also still trails Opus 5 on HLE, Terminal-Bench 4.0, and similar benchmarks. Wait for third-party verification before drawing conclusions.
Source: Read the original. This article is a summarized digest for reference.
