Report: Anthropic downgrades Opus on training-adjacent jobs
A researcher says fine-tuning work got Opus 5.5 cut to Opus 5 — a taste of the rewritten usage policy landing Nov 12, which also bans proxying Claude.
Copy markdown
What reportedly trips the downgrade
The researcher says work touching RFT, SFT, or training larger models now gets Opus 5.5 silently downgraded to Opus 5 — or the session killed — even when the rollouts came from an open model and Claude was only orchestrating. Anthropic hasn't confirmed it, and a Hacker News thread on the report drew close to 190 points.
Proxying Claude through consumer plans ends Nov 12
The rewritten usage policy — published Oct 8, effective Nov 12 — bans accessing Claude through 'unauthorized means,' explicitly naming services that route requests through consumer subscriptions or circumvent sign-up and login. If your app or bot rides on someone's Pro or Max seat, that path closes; move to the API.
You're now liable for whatever your agents do
The new policy makes you explicitly responsible for ensuring agents you build or deploy comply — including actions taken through tools, browsers, and connected systems. And enforcement now triggers when Anthropic 'suspects you may have violated' (up from 'learns that you have'), with new 'warn' and 'limit' steps short of an outright ban.
Apple Intelligence opens to China devs on Qwen Nov 4
Apple's China build of Apple Intelligence — running on Alibaba's Qwen across iOS, iPadOS and macOS — reportedly opens a developer preview on Nov 4, ahead of a December consumer rollout with iOS 27.3. For anyone shipping to the China App Store, it's the first on-device Apple AI stack you can target there, and Alibaba says user prompts and responses won't be used to train Qwen.
Five small Qwen models reportedly beat Fable 5 on hard coding
An open 'memdoor' workflow runs five Qwen3.8-27B models as a planner/ideator/solver/manager team that talks only through files on disk, and reportedly solved 49/50 LiveCodeBench Hard problems — mean 9.8/10 vs 8.0 for Claude Fable 5 solo — for $35.79 across 15.3M output tokens. Clonable this weekend, but Hacker News flagged that the shared files may leak solutions between agents, so re-run it isolated before trusting the score.