Anthropic reverses course: Fable 5 stays in Max and Team Premium plans
Anthropic has walked back its decision to pull Fable 5 from subscription accounts: from July 20, Fable 5 is included in all Max ($100–$200/month) and Team Premium plans at 50% of normal usage limits. Pro and Team Standard users keep access through usage credits and get a one-time $100 credit; the $20/month plan remains excluded. Simon Willison reads the reversal as competitive pressure from GPT-5.6 Sol and Kimi K3 — a subscription without the vendor's best model was getting hard to justify — and notes the open question of whether serving Fable 5 to subscribers forces Anthropic to divert GPU capacity from training.
Sources & depth
Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?
Charles Azam benchmarked Claude Fable 5 against GPT-5.6 Sol on KIRO, an unpublished NP-hard fiber-network optimization problem with a search space around 10^1223, across matched 30-minute agentic runs. Plain runs: Fable 5 averaged 32,386 points (tight 319-point range) vs Sol's 34,261 (lower is better; Sol's spread was 6x wider). The twist is the /goal persistence feature: it won 4 of 6 trials head-to-head yet made *mean* scores ~750–870 points worse for both models — on optimization tasks, persistence just amplifies whatever solver strategy the model committed to, good or bad. A clean demonstration that win-rate can mask mean regression when evaluating agent features.
Sources & depth
GPT-5.6 used a prompt to close a 30-year gap in convex optimization
In the wake of OpenAI's Cycle Double Cover conjecture proof announcement, a researcher reports GPT-5.6 (Sol Pro) proved a 30-year-open lower bound in convex optimization: that optimizing convex Lipschitz functions requires Ω(d²) function evaluations, matching the complexity of a decades-old algorithm. The catch, per the discussion: it took a 10-page prompt distilling a year of the author's prior attempts with GPT-5.4/5.5 — including the function class that ultimately worked — so commenters split on whether this is AI capability or human-directed formalization ("a year plus 148 minutes"). The proof has not yet been independently verified. Reddit thread unfetchable from the desk — gist grounded in the HN discussion below.
Sources & depth
The Kimi K3 Moment
Stephen Bochinski argues Chinese open-weights models have hit their "DeepSeek moment, but for real this time": he claims Kimi K3 matches Claude's coding output quality at $3/$15 per million tokens against Claude's $10/$50, with a $39/month coding plan undercutting Anthropic's metered tiers, and cites GLM 5.2 (MIT-licensed) beating Claude on Semgrep's cyber benchmarks. His policy read: US capability restrictions handed the unrestricted-model market to Chinese labs, and he predicts an automotive-style protected-but-uncompetitive domestic industry. An opinion essay — the parity and pricing claims are the author's, not independently measured here.
Sources & depth
Setting up your spare Mac for Claude Code to control, a step-by-step guide
A 16-step guide to turning a spare Mac into an always-on machine Claude Code fully controls: fresh isolated account with no Apple ID, SSH with passwordless sudo, a persistent tmux server for screen capture and input control, optional Tailscale/VPN for remote access. The isolation is the point — the agent gets broad system permissions on a machine with nothing sensitive on it. What's new versus container sandboxes is full GUI access: Claude can drive native Mac apps (the author demos Unity) rather than just a shell.
Sources & depth