Mozilla’s State of Open Source AI v1.1 (Sept 15) pegs the open–closed gap at about 4.4 months on METR task horizons. Top open models are far cheaper—e.g. Kimi K3 ~2 points behind Claude Fable 5 at roughly 30% the price—so Mozilla advises open as the default for most work.
New CLAUDE CODE SUBAGENT EFFORT sets the effort level subagents use when nothing else names one, and CLAUDE CODE DISABLE ULTRATHINK switches off the ultrathink keyword. Fast mode for Opus 4.8 is deprecated and will error from Nov 30, 2026 with no fallback, so move fast mode work to Opus 5.5…
Run gh actions lock to write .github/workflows/actions.lock , which pins every action your workflows use (transitive ones included) to an exact commit even if the tag moves, and Dependabot now updates the lock in the same PR. After a preview limited to Enterprise Cloud, it's out on Free, Team and Enterprise, with a Require lockfile policy and a stale lockfile check you can run in CI…
Released Oct 30 under Apache 2.0, the 9B text and image model reuses a block of 24 decoder layers twice with shared weights, giving about 14.6B parameters' worth of compute per pass, with 262K token training context aimed at long context coding and tool use agents. It's a preview: loading needs trust remote code=True (about 38 GB of BF16 weights) and there's no technical report yet, so treat its benchmark chart as unverified (GitHub)…
New this week under MIT, npx skills add aicodeking/deepseek stack installs SKILL.md workflows such as ds plan loop , ds verify loop , ds review loop and ds best of n that force a DeepSeek powered agent to check each step against real command output, and it works in Claude Code, Codex, Cline, OpenCode and other harnesses. You can also install a single skill with skill , and each one is short enough to read before you let it run (skills.sh listing)…
Released Oct 30 under Apache 2.0, the 9.5 GB model uses one policy to pick, turn by turn, between screenshot grounded GUI actions and a persistent shell, and Microsoft reports 66.0% on OSWorld Verified with a 50 step budget, ahead of some larger open agents. Those scores are Microsoft's own so far, it needs trust remote code=True or the provided SGLang container, and since it can run bash , keep it in a sandbox (project page)…
Spotted something we missed? Start a thread.
