The Pulse – what changed in AI · 20.10.2026

Changes from October 17–20, 2026 for people who build with AI. Each item links to its primary source.

DeepSeek ships V4.1-Flash-1017 with open weights

Models · 2026-10-17

A refresh of V4.1-Flash, live on DeepSeek's site, app and API, with weights on Hugging Face and ModelScope. DeepSeek says it scores higher, finishes tasks faster and writes shorter outputs; the benchmark figures are its own.

Source: DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real

DeepSeek opens Search and Fetch APIs, free in Claude Code and Codex until November

Harnesses · 2026-10-17

Search and Fetch cost $0.75 per 1,000 successful requests each. Calls made through DeepSeek Harness, Claude Code or Codex are free until billing starts on November 1 (Beijing time).

Source: DeepSeek API Docs: Search and Fetch

MiniMax M3.1 Flash: a cheap agent model at $0.15 / $0.60 per million tokens

Models · 2026-10-19

MiniMax says the new model runs nearly 3× faster and cuts the cost of agentic tasks by 70% compared with its predecessor (vendor figures). It is available now through the MiniMax API; the weights are promised for later.

Source: MiniMax M3.1 Flash: Redefining Agent Performance and Cost

Claude Code 2.1.304 adds guided auto-mode setup and a usage warning on resume

Harnesses · 2026-10-19

New /auto-mode-setup teaches auto mode about your environment, including Bedrock and Vertex. Pro and Max users now get a warning when resuming an idle conversation would take a noticeable share of the 5-hour limit, and Code Review owners can set Quick, Standard or Deep levels per repo.

Source: Claude Code changelog: 2.1.304

---

New to building with AI? Start with the free first lesson: https://www.vibe-coding.academy/start