Muse builds mini-apps from one sentence; agent-run costs drop; AI spend stalls

🔥The one thing to know

🧱New ground broken

• Claude Fable 5.1 — costs about 25% less than its predecessor for typical work, and up to roughly 45% less for agentic work, because cache reads — an agent re-reading context it already processed — drop to $0.25 per million tokens. Input is $10 and output $50 per million tokens. Cache reads are the line item that decides whether an always-on agent feature survives its first cost review — rerun the unit economics before rejecting one on price. A new Enterprise Frontier Safeguards setup adds zero-data-retention privacy with data stored in customer-controlled infrastructure, phasing in for enterprises later this fall. Anthropic

• Alexa shopping by text — invited U.S. customers can text Alexa's agent to search Amazon, add items to a cart, and buy in plain language; it also tracks orders, compares products, and finds similar listings from a shared photo. Amazon built its own shopping agent two months after blocking Meta's Muse from its site — platforms are keeping the agent surface in-house, which is the pattern to plan around. Uninvited users can join a waitlist; broader rollout is due in the coming weeks, ahead of the holidays. TechCrunch

🚀Agent features and products that impact product management

• Cursor voice chat — You can now hold a live voice call with a Cursor agent — describe a change out loud and hear short progress notes while it works — in the IDE, the Agents Window, on the web, and on iOS, where Siri can start the call. Agent supervision stops being a reading task; expect your team to check in on runs by voice and skim fewer diffs. It runs on Grok Voice with several voices to pick from, and with Privacy Mode on there is zero data retention. Cursor

• Agentic Identity Broker — an MIT-licensed broker from Zalando Engineering that lets AI agents act for a user across Google, GitHub, Databricks, Linear and other services with scoped, revocable permissions instead of shared credentials — it handles consent, delegation chains and a secure token vault, in the call path via gateways like Agentgateway. This is the plumbing your roadmap needs the moment an agent touches a user's other accounts, and it is already proven. Zalando uses it in production for several internal MCP servers; the roadmap adds agent-to-agent token exchange and user-facing activity logs. GitHub (Zalando Engineering)

🧭How product teams work now

• Ask HN agent setups — Three readers counted every comment in a 663-comment Ask HN thread: Claude Code leads with 180 of 424 tool-users, Codex follows at 103, and 156 people run two or more coding agents. The habits that repeat: parallel agents each in their own git worktree, a fresh second agent reviewing every diff, and strong models planning while cheap models build. Multi-vendor, multi-agent with a review loop is now the mainstream setup — a fair benchmark for your team's tooling budget. One commenter runs Claude for planning, Codex for review and DeepSeek for building, under $100 a month. The AI News Report

  “all you need are Git worktrees and a decent OS environment.”

📰Market and platform news

• Ramp spend data — Ramp's card data shows AI token spend grew just 1.5% across August and September, against 42% in July, even as volumes rose — vendor price cuts and thriftier customers are capping revenue. Anthropic said its Fable 5.1 and Opus 5.5 price cuts alone cut revenue growth by about $600 million. Assume your model costs fall every quarter, but so does everyone's margin on them — it is a buyer's market for tokens. October spend rebounded 8.9%; Bloomberg reports Anthropic's valuation talk sank to $1.2–1.3 trillion from $2 trillion. TechCrunch

• Inkling 2 Small — Mira Murati's Thinking Machines Lab released a 594B-parameter mixture-of-experts model, 17B active per request, that Artificial Analysis rates the strongest open model outside China — a title conditional on the weights actually landing, scheduled for Apache 2.0 release next week after safety testing. A near-frontier open-weight option keeps pressure on your model costs and gives agents a self-hostable fallback. Unconfirmed: Business Insider reports a $5 billion raise at a $40 billion valuation, led by Accel. Thinking Machines Lab

• Clutch Security — Palo Alto Networks is in reported talks to buy Clutch Security, which controls what AI agents can touch inside company systems, for $500 million in cash plus milestones — about 50 times its approaching $10 million in annual recurring revenue. Peers went similarly: Cyera paid $1 billion for Oasis Security, Cisco about $400 million for Astrix. Agent permissions are becoming a line item enterprises buy off the shelf; expect the build-vs-buy question on your own roadmap. It would be Palo Alto Networks' seventh acquisition this year, per PitchBook. TechCrunch

• Meta scam-ad documents — Reuters-reported internal Meta documents put expected revenue from scam and prohibited-product ads at roughly $16 billion, about 10% of 2024 revenue, with users seeing some 15 billion "higher-risk" ads a day. Mark Cuban says the platform keeps running AI deepfake scam ads built with his face and voice. Trust in AI-generated content is now a revenue-and-regulation problem — the SEC is looking into Meta's ad practices — so budget for deepfake policing if your product carries ads or user media. Meta spokesman Andy Stone calls the documents a selective view and says the company removed more than 134 million scam ads in 2025. Yahoo Finance

🌍The big AI stories today

• OpenSSH 10.6p2 — shipped a critical workaround and will release more often, because AI-generated security reports keep arriving — and adversaries can find the same bugs. Hachyderm (Damien Miller)

• Google Wyoming data center — Google announced a $17B Wyoming data center with carbon-capture gas power — two days after voters rejected data centers in 40 of 41 ballot measures. Google

💡One takeaway