Moonshot published Kimi K3.1 Pro's full weights, and its own benchmark table puts the model 0.7 points behind Claude Opus 5.5 on Terminal-Bench 4.0 and ahead on SWE-Marathon, though that test ran mostly through Claude Code. Four more stories inside.
Today's lead is a benchmark fight: Moonshot's newly open Kimi K3.1 Pro lands within a point of Claude Opus 5.5 on one coding benchmark, by Moonshot's own numbers. A smaller rival claims a narrower title the same day. Also today: Claude Code spends far less writing status lines for its background agents, a cost an Anthropic engineer called "a lot for a status line" last week. Conductor, which runs coding agents in the cloud, says it's joining Meta and shuts its cloud service Nov 20. And Claude Code's product lead since its creation left Anthropic for OpenAI, two weeks before Anthropic's planned IPO.
What changed this week #
Moonshot says Kimi K3.1 Pro comes within a point of Claude on coding
Moonshot AI published the full open weights for Kimi K3.1 Pro, a 2.8-trillion-parameter model it had only sold through its API since October 23. On Moonshot's own benchmark table, it scores 65.7 on Terminal-Bench 4.0 against Claude Opus 5.5's 66.4, a gap of 0.7 points, and it beats Opus on SWE-Marathon, a long-horizon coding test, 51.7 to 46.3. Artificial Analysis, running its own test, scored K3.1 Pro even higher: 67.2%, the top result in its comparison. The lead flips on Terminal-Bench Science, where K3.1 Pro scores 38.1 against Opus's 63.3. One catch: Moonshot ran K3.1 Pro mainly through Claude Code itself, so the number measures a model-and-harness pair, not the model alone. The full weight set runs about 5.6 terabytes.
Ricursive's $32M Ripple-65B claims the top US open coding score
Ricursive Intelligence released Ripple-65B Preview, scoring 80.52 on Terminal-Bench 2.1, an older version not comparable to Kimi K3.1 Pro's test above, but still the top US open-weight score. It cost $32M to train and ships under Apache 2.0 at $0.60 per million output tokens, per VentureBeat. The "US" qualifier does real work: Kimi K3.1 Pro, out the same day, is Chinese. One commenter called Ripple "SOTA in the same way Tesla are SOTA EVs - the best you're permitted to buy in the USA."
Anthropic cuts Claude Code's background-subagent status costs by 99%
Claude Code was re-reading each background subagent's whole conversation every 30 seconds just to write a one-line status, a cost we covered Oct 29 when an Anthropic engineer called it "a lot for a status line" and promised a cheaper mode. v2.1.316 ships it: Haiku 5.5 now writes the status from short running notes, 99% cheaper per refresh, so affected sessions use about 5% less of a plan's limit. One gap: no first-party Anthropic page, only a write-up and two social posts.
Conductor, a cloud coding-agent orchestrator, is joining Meta
Conductor, whose app runs Claude Code, Codex, Cursor and OpenCode agents in cloud sandboxes, says its team and technology are "joining Meta," the wording it uses instead of "acquired" or "acqui-hired." No price was disclosed. Conductor Cloud goes read-only Nov 13 and shuts down Nov 20; the desktop app keeps working, with new models and bug fixes, "until early 2027," while Conductor builds "something new for you to use early next year," unnamed. No Meta statement turned up.
Cat Wu, Claude Code's product lead, leaves Anthropic for OpenAI
Cat Wu, who took Claude Code's product from an internal prototype to what Business Insider calls "a commercial juggernaut," left Anthropic last week for OpenAI, to lead product management for its "core products" team under Codex lead Tibo Sottiaux. Not her first defection: she and Claude Code's creator, Boris Cherny, left for Cursor in 2025 before boomeranging back within the year. Anthropic declined to comment. OpenAI confirmed the hire. It lands two weeks before Anthropic's planned IPO.
Also worth your time #
- Runway ships Dev plugins wiring its video/image/audio APIs into Claude Code, Codex and Cursor
- GitHub Copilot reaches Xcode's own coding assistant in a public preview, per GitHub's changelog
- Garry Tan tells Walter Isaacson he spends up to $200K a year on tokens and shares 23 Claude Code skills on GitHub
- OpenAI's Greg Brockman says agentic coding tools helped drive a Q3 "revenue inflection," per the WSJ via Techmeme
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.