DeepSeek Releases V4.1-Flash-1017 for AI Agent Tasks

DeepSeek, the Chinese AI lab, released a new model, V4.1-Flash-1017. DeepSeek says that in its own benchmarks, the model beats four rivals. Those are Anthropic's Claude Opus 5, OpenAI's GPT-5.6-Sol, Moonshot AI's Kimi K3, and Zhipu AI's GLM-5.3. The company says its model wins at coding tasks and at agentic tasks, where the AI uses tools on its own to get the job done. On the DeepSWE benchmark, the new version scored 77 percent accuracy at a cost of around 20 cents per task. The previous version scored 74.2 percent at around 50 cents per task. Through Claude Code and Codex, two tools developers use to run AI coding agents, DeepSeek is keeping these capabilities free until October 31 at 16:00 UTC, Coordinated Universal Time. After that date, the company will charge $0.75 per 1,000 search queries.

These numbers come only from DeepSeek's own testing, so they stay unverified until independent researchers confirm them. Chinese AI labs, including the makers of Kimi K3 and GLM-5.3, have already published bold self-reported results of their own, and real-world work with AI agents has often turned up problems that no benchmark catches. Free access through Claude Code and Codex through the end of October gives developers a risk-free window to test the model on real tasks. Whether it can replace Claude Opus or Fable, the models used for the hardest agentic work, is a separate question these benchmarks do not answer.