DeepSeek's Oct 17 update to V4.1-Flash (763B params, 1M context, MIT license) raises Terminal-Bench 4.0 from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0 in its own tests, while finishing tasks faster with shorter outputs. It's on DeepSeek's API and app, plus Vercel AI Gateway as deepseek/deepseek-v4.1-flash-1017.
As of Oct 17, any DeepSeek API key can call POST /v1/tools/search (ranked web results with quality and authority scores) and POST /v1/tools/fetch (one page as Markdown), each billed only on successful calls and capped at 5 requests per second per account. Server side search and fetch from DeepSeek Harness, Claude Code, and Codex on the DeepSeek API stays free until Oct 31, 2026, 16:00 UTC…
goose v1.55.0 (Oct 16) propagates W3C trace context from ACP prompts into MCP tool calls, surfaces provider x request id on responses and errors, and binds subagents to their parent's hooks, mode, and approvals. It also caps auto compaction at 225k tokens by default and prefers OpenAI's Responses API for reasoning models…
In @open multi agent/core 1.5.0 through 1.21.2, an agent's file write could create files outside its working directory by writing through a dangling symlink; v1.21.3 (Oct 16) now resolves those links one path component at a time and rejects any that point outside the root. Upgrading needs no code changes, and every built in filesystem tool shares the fixed check…
As of Oct 16, adding https://api.heypcb.ai/mcp to Claude Code, Codex, Cursor, or ChatGPT lets agents draw schematics, place parts, route boards, run checks, export KiCad projects, and design enclosures. In HeyPCB's demo, Claude Code laid out a 145 part 8x8 RGB display board, fixed a rotated LED that wouldn't route, and finished with zero DRC and ERC errors in KiCad…
OpenAI's Oct 16 release notes say Slides lets you build decks from a chat, files, or an outline in ChatGPT Work, co edit them in Space, and export to PowerPoint, PDF, or Google Slides on the web. GPT 6 Pro in Chat also now splits a task across multiple agents running in parallel, each with a viewable record of what it did, rolling out to eligible Pro users…
Spotted something we missed? Start a thread.