Update: Memdoor open-sources the 5-Qwen team that beat Fable 5
A single-binary Apache-2.0 agent, ~$0.72 a run. Also: Claude Code's ultrathink kill-switch and pooled team budgets, Kimi K3.1 weights due Nov 4.
Copy markdown
Five small open models, one YAML workflow
Memdoor is a single-binary, Apache-2.0 CLI agent (bring-your-own-keys, no Docker or Python) whose published team-solve workflow runs five Qwen3.8-27B agents — planner, two ideators, solver, manager — that coordinate only through files on disk. It scored a mean 9.8/10 against Fable 5's 8.0 on the first ten LiveCodeBench Hard problems, with all 50 runs public at roughly $0.39 and ~40 min each — though it's a ten-problem sample on possibly-contaminated early-2025 contests, using the project's own numbers.
Claude Code adds an ultrathink kill-switch
Claude Code v2.1.316 ships CLAUDE_CODE_DISABLE_ULTRATHINK to turn off the 'ultrathink' trigger word, plus pooled group budgets — a pooled_budget field in the spend-limits admin API that lets a team share one spend cap instead of per-seat limits.
Kimi K3.1 Pro weights land by Nov 4
Moonshot's K3.1 Pro is live on the API now at $2/$10 per million tokens with a 1M-token context, and open weights are promised by Nov 4 — a near-frontier coder self-hosters can download within days.
OpenCode ships three releases in a day
OpenCode v2.2.3–2.2.5 add git worktree and base-branch shortcuts, a composer shell-mode toggle, and desktop terminals that survive server restarts — on top of the Claude Code /connect login that landed this week.