Fable 5.2 ships 20% cheaper, Claude Code leaks secrets to GitHub, and model half-life hits six weeks

🔥The one thing to know

🧱New ground broken

• Fable 5.2 one-shot game — is what a practitioner woke up to after handing Fable 5.2 his nephew's hand-drawn design documents before bed. The agent threw out the old codebase, generated about 175 images and 140 audio clips, and even wrote a bot that played all twelve stages in a headless browser to check for problems. This is a shipped, playable artifact, not a lab demo — a full multi-level game is now a weekend-scale build, which resets what a 'prototype' costs to promise a stakeholder. The run took roughly 600 Claude requests over two days, around 340M tokens, most of it cache reads. NotesByLex

• Skills in Google Workspace — are shareable instruction files, built on the open SKILL.md format, that teach Gemini a team's rules, templates and reference documents — now rolling out across Docs, Gmail, Slides, Drive and Chat, where Gemini can create and apply them mid-task. Encoding your product rules in files an assistant can read just became a default for non-technical teams; a PM can hand sales or legal a skill instead of a walkthrough. English only for now, missing from Sheets and Meet; Rapid Release domains started getting it November 9, Scheduled Release domains from November 30. Google Workspace Updates

• Muse Ember — is Meta's smallest model, aimed at the calls agents make all day — tool use, routing, extraction, small coding fixes — and designed to pair with a bigger planner model in the same harness. Execution at cents-per-task is what makes always-on agent features affordable at consumer scale; if your roadmap has an agent feature parked on cost grounds, re-run the numbers. List pricing is a dime per million tokens in and half a dollar out, with a 1M-token context window; Meta's own figures put a typical scored task at about $0.14. Meta

🚀Agent features and products that impact product management

• Claude Code settings leak — Security researchers at Apiiro found Claude Code's local settings file — which stores the exact commands users approve, sometimes with tokens and passwords embedded — on public GitHub in over 100,000 cases; roughly 1,100 files across 63 organizations contained real secrets. Your team's coding agent is now a credential-leak vector with a bug trail going back to August 2025; a repo-level check for the settings file belongs in this sprint's security review, not next quarter's. 215 of the exposed secrets slipped past both gitleaks and TruffleHog, and Claude Code's lead says the team is working on prevention. Ars Technica

  “looking into what we can do to prevent this”

• Medium TK — is Medium's new writing app with an assistant summoned by typing @TK inside the editor, shipping with skills for editorial and headline feedback, plus optional notes authors publish explaining how AI was used in a piece. Provenance labeling in a mainstream publishing product is the pattern to watch if your roadmap touches AI-assisted content — and its MCP server means a user's own agent can plug into the workflow. Free for anyone with a Medium account, web only at launch, with a desktop app on the roadmap. TechCrunch

  “make the provenance of writing clear, transparent, and inspectable, right down to the prompting and source material.”

• Instagram Effect Builder — is a chatbot that builds custom Story filters from a text prompt and optional style reference images, lets users preview on their own photos and iterate in follow-up prompts, then submit effects for others to use. User-generated agentic content with a human safety gate and attribution back to the creator is a template worth stealing for any creative-tool roadmap. Live now on the web for US users over 18; every submitted effect clears a safety review before other people can use it. The Verge

🧭How product teams work now

• How I AI: William Imoh — is a Lenny's Podcast episode where the founder of Hackmamba describes building Mural, an AI language tutor, with Codex writing the Swift and Kotlin despite him knowing neither — and notes that as models improved, most of his carefully built agent skill library became unnecessary. Two lessons for PMs: he structures requests as epics and user stories tied to user problems, and he found heavy prompt scaffolding decays within months — spec quality is the durable skill. He prices a 30-minute real-time GPT-Live tutoring session at roughly $3 to $5 — steep for software, still cheap next to a human tutor. Lenny's Newsletter

• Ask HN: agent context — is a Hacker News discussion (86 points, 51 comments) where people running agents for multiple clients describe real setups: a gatekeeper script that refuses to mark work done until the project wiki is updated, per-client directories with separate knowledge bases, and plain Markdown indexes searched semantically. The working answer to agents forgetting decisions is durable, human-readable notes plus mechanical checks — a pattern your team can adopt this week without buying anything. One widely endorsed approach, Johnny.Decimal, claims about half an hour to set up, with CLI tools and agent skills included. Hacker News

📰Market and platform news

• Vercel AI Gateway Production Index — is Vercel's monthly read on production AI traffic, and November's shows breakneck churn: in October the majority of text-model calls went to releases younger than 60 days, and a model that peaks in its first month sheds about three quarters of its peak teams within another month. Pinning a roadmap or contract to a named model is now a losing bet — the durable decision is a swappable model layer. New teams in October chose Anthropic over OpenAI nearly two to one, and voice adoption among paid teams climbed from 4% in July to 11%. Vercel

• Gemini connected-apps review — is what a GitHub repository documents, with reproduction steps: Google has reversed its 2023 promise that personal Workspace content stays away from human eyes, and casual conversation — no @ tags — can probabilistically pull full unredacted emails into a session. If your product connects user accounts to an assistant, 'a person may see this' is now the honest disclosure; design for it before an enterprise buyer does. The author's tests produced four distinct outcomes, ranging from no retrieval to a complete email dump including headers. GitHub

• Apple Intelligence Usage Packs — are $50 add-ons that buy extra AI capacity when a user's monthly allowance runs out, arriving first in Apple's pro creative apps under the Creator Studio banner. A metered top-up — not a flat subscription — is a fresh pricing shape for capped AI features; worth a look if your own usage-based pricing punishes your heaviest users. It's the first concrete paid mechanism Apple has actually shipped for heavier AI consumption, after signalling it intended to monetize beyond-the-cap use. 9to5Mac via Techmeme

🌍The big AI stories today

• Anthropic IPO — Reuters saw emails showing Anthropic circulated and then walked back IPO pricing details late Monday — treat anything you've planned around its listing timing as soft. Reuters via Techmeme

• Nvidia investigation — The probe follows Huang's admission of an affair with a woman US officials reportedly believe is a Chinese intelligence asset, with House chairs seeking a closed-door meeting November 18 — more export-control and supply noise for anyone building on Nvidia silicon. Business Insider

• TPU 8tx and 8i Max — One is built for confidential inference on sensitive data, the other for huge context windows, and Google open-sourced TorchTPU so PyTorch code runs natively — relevant if you're weighing inference lock-in for agent workloads. Google

💡One takeaway