The vendors are starting to eat the cost of a bad answer

The 60-second catch-up

The bill is moving from attempts to outcomes. OpenAI has begun letting select major accounts pay only when the work lands, according to a report The Next Web says it could not independently verify. The named prices around it are firmer: Intercom charges $0.99 for each conversation its Fin agent resolves and nothing for the ones it does not; Zendesk bills only "Verified Resolutions" confirmed by an LLM check within 72 hours, at roughly $1.20–$1.50 on committed volume; Salesforce launched Agentforce at $2 per 24-hour session whether or not anything was resolved, and customers said so. The pressure behind all of it is token billing — one developer running a hundred agents in parallel accumulated $1.3m of OpenAI tokens in thirty days, because cost scales with attempts, not results. Outcome pricing moves the cost of a bad answer from the buyer to the vendor. That is why it works for support, where a resolution is a field in a database, and stalls everywhere else.

→ do this: before you price an agent, write down the single outcome you would be willing to bill for. If you cannot define it in one sentence, you are selling attempts.

Access, not price, is becoming the scarce thing at the frontier. Tomasz Tunguz makes the case plainly: governments are treating frontier models as sovereign infrastructure, and the gate is no longer cost but permission. Z.ai shipped GLM-5.3-Flash, a 320B model, under a plain MIT licence on 26 August — then put its GLM-5.3 flagship out two days later under a bespoke licence requiring any model-as-a-service host with more than $10bn of revenue to pass a security review first. OpenAI served Cursor a termination notice effective 12 November. Nvidia, meanwhile, has committed $26bn over five years to open-weight development, plus $13bn for Hugging Face and $7bn for Poolside — open weights as structural defence against its own cloud customers. Meta's new Muse Code CLI is free to install; the model behind it is the toll booth.

→ do this: for every model you depend on, ask whether your access can be revoked, not just repriced. Keep one open-weights path warm enough to switch to.

Consumer AI just acquired a compliance department. The European Commission designated ChatGPT, Reddit and Roblox very large online platforms under the Digital Services Act, after each crossed 45 million monthly users in the EU. That brings obligations on illegal content and minors' privacy, fines of up to 6% of global revenue, and a deadline of end-December. In the same week Instagram said it will demote AI-generated influencer accounts that do not clearly label themselves. Both moves point the same way: scale in the EU now arrives with a filing cabinet attached, and labelling is moving from ethics to product spec.

→ do this: if you have EU users and anything resembling a feed, put provenance labelling on the roadmap this quarter rather than the one after.

Tools

Build & ship

• ⭐ Muse Code — Meta's multi-agent coding agent for the terminal, built for Muse Spark · Free CLI · pay-per-token API
• diffium-db — Live TUI diff of what an agent or migration is doing to your database, structure and rows · Open source (MIT)
• x1 — Describe an iPhone app, get a plan, screens and a testable build · Free to start, no card

AI & agents

• slotstream — Runs a 125B mixture-of-experts model (104 GB at 4-bit) on a 48 GB Apple Silicon Mac · Open source (MIT)
• BrowserAct — Reusable cloud web scrapers from a prompt, or a browser for your local agent · Free trial, then credits

Design & create

• ⭐ Ambient CSS — Physics-based lighting for CSS — define a light source and every shadow and highlight follows, calibrated against Blender raytraces · Open source (MIT)
• before.click — A browsable gallery of App Store screenshot sets, by category · Free

Growth & ops

• Weedout — Strips AI-generated filler out of your YouTube feed · $1.99 one-time (macOS 13+)
• Movie Scene Map — 15,565 real filming locations across 166 countries; data downloadable as CC0, with a read-only MCP endpoint · Free
• HN Match — Matches you against the current Who's Hiring thread by skill · Free

Resources

Steal the template

• An end-to-end process for getting real design out of AI — Anshu Chimala, who led design and engineering R&D at Apple for twelve years, on why models default to the same purple-gradient landing page and how to break them out of it. The concrete trick is String Seed of Thought: have the model generate a random alphanumeric string with a shell script and derive the creative direction from it, so variety comes from outside the model rather than from asking politely for it. Free preview, paid in full.
• Workflow Lab: moving between design and code with agents — Figma's own account of the hand-off, from the side that owns the file format.
• Agent memory as a file format — Cal Paterson proposes a portable file format for agent memory instead of another vendor's database.

Learn the craft

• Agency and Agents — Ethan Mollick on what actually changes when the software has initiative.
• The Custodial Era of UX: cleaning up after AI — Nielsen Norman Group naming the debt that accrues when generated interfaces ship faster than anyone can review them.
• A deep dive into ZCode — Flavio Copes, long and unhurried.
• Design's legibility gap — Peter Merholz on why design work keeps being illegible to the people funding it.
• Two Adobe AI features worth a web designer's attention — narrow, practical, no hype.

Benchmarks you can re-run

• Open Design System Bench — Christoph Hellmuth's method for measuring how AI-ready your design system actually is, rather than asserting it.
• The efficient frontier of LLM inference — Baseten plots the latency/throughput/cost trade-off you are actually choosing between.
• How accurate have Ed Zitron's AI skeptic predictions been? — Dan Luu scores a prominent sceptic's calls against what happened. Useful whichever side you started on.
• TimesFM-3 — Google's zero-shot multivariate forecasting model, with GIFT-Eval and FEV-Bench results, weights on Hugging Face and code on GitHub.

Reads

☕ Under 5 minutes

• The price of entry to the frontier — Tomasz Tunguz. The argument the catch-up above is built on, with every claim footnoted.
• Codex bundles a full copy of LibreOffice — Simon Willison, on what is actually inside the app you installed.
• OpenAI has started letting some customers pay only when the AI works — the outcome-pricing report, with the comparative vendor prices laid out.
• Google is building AI Rooms for Gemini Enterprise — shared agent workspaces, spotted before the announcement.

🍵 5–10 minutes

• diffium-db: a live diff of what agents do to your database — Denislav Gavrilov on the design choice that makes it useful: when it cannot make a precise claim about a table, it says less rather than more.
• Anthropic banned me for "suspicious signals" — a first-person account of automated account termination with no route to a human. Read it as a product risk, not a grievance.
• FBI probes service selling 153M+ driver's licenses — Krebs on Security.
• "Zlibrary my beloved": staff chats extolling piracy cited in Sony suit — what internal Slack looks like in discovery.
• Instagram will demote AI-generated influencers who don't label themselves

📚 Longer, worth it

• Introducing Solaris — Runway's "Interface World Model": generated software you can interact with in real time at 720p, built on its Gen-4.5 video model. The reconstruction benchmark asking what is lost when an interface is translated into code is the interesting part.
• Atlas: a world model for spatial intelligence — World Labs.
• Path to Astra — OpenAI setting out critical capabilities and the safeguards it says go with them.
• My local model setup on an M4 Pro Mac mini — Kevin Lewis, specific about what works and what does not.
• Fine, I'll build my own text editor — David Bushell.

Numbers

• 45 million — monthly users in the EU, the Digital Services Act threshold that ChatGPT, Reddit and Roblox each crossed (European Commission).
• 6% — of global revenue, the maximum fine the three now face if they miss the end-December compliance deadline (European Commission).
• $0.99 — what Intercom charges per conversation its Fin agent resolves, and nothing for the ones it does not (Intercom).
• $1.3 million — OpenAI tokens one developer accumulated in thirty days running a hundred agents in parallel (The Next Web).
• 320 billion — parameters in GLM-5.3-Flash, released under a plain MIT licence, two days before its sibling flagship shipped under a restrictive one (Z.ai).
• $26 billion — Nvidia's five-year commitment to open-weight model development, on top of $13bn for Hugging Face and $7bn for Poolside (WIRED).
• 75% — the fall in Alibaba's quarterly earnings reported the week before its $10bn placement, driven by AI capital expenditure (Reuters).
• 5% — OpenAI's share of Cursor's traffic at the point OpenAI served notice of termination (Cursor, Business Insider).
• 15,565 — real filming locations in a free CC0 dataset that also answers queries over a read-only MCP endpoint (Movie Scene Map).
• $10 billion — Alibaba's share placement, the largest primary follow-on offering ever by a Hong Kong-listed company, announced a day before its Wan3.0 video model shipped (Reuters).
• 153 million — driver's licenses offered by the identity-verification service now under FBI investigation (Krebs on Security).

What happened

Models & math

• Runway introduced Solaris, an "Interface World Model" that generates interactive software in real time at 720p, adapted from its Gen-4.5 video model. Access is by request only for now (Runway).
• Google released TimesFM-3, a zero-shot foundation model for multivariate time-series forecasting, evaluated on GIFT-Eval and FEV-Bench, with weights on Hugging Face (Google Research).
• World Labs published Atlas, a world model aimed at spatial intelligence (World Labs).
• Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 (Anthropic).
• OpenAI published Path to Astra, its statement of critical capabilities and frontier safeguards (OpenAI).
• Z.ai shipped GLM-5.3-Flash, a 320B model, under a plain MIT licence on 26 August, then released its GLM-5.3 flagship two days later under a bespoke licence requiring hosts above $10bn revenue to pass a security review (Z.ai, The New Stack).

Safety & policy

• The European Commission designated ChatGPT, Reddit and Roblox very large online platforms under the Digital Services Act, each having passed 45 million monthly EU users. EU tech chief Henna Virkkunen said the three "will now be held to a higher standard of scrutiny and accountability." They have until the end of December (European Commission).
• Instagram will demote AI-generated influencer accounts that do not clearly label themselves (Instagram).
• Internal staff chats praising Z-Library were cited in Sony's copyright suit against Anthropic (Ars Technica).
• The FBI is investigating a service selling access to more than 153 million driver's licenses (Krebs on Security).
• The Department of War launched OpenAI's ChatGPT Mil and Starshield AI's Grok for Government on GenAI.mil, both accredited for Controlled Unclassified Information at Impact Level 5 and built to serve more than 3 million personnel. The platform has onboarded 1.7 million unique users in nine months (Department of War).

Business moved

• Alibaba launched its Wan3.0 video model, which generates 30-second videos from documents, spreadsheets, slides and web pages, a day after a $10bn share placement to fund AI spending. It had reported a 75% fall in quarterly earnings the week before (Reuters).
• Salesforce and Anthropic launched Claudeforce, making Claude the default across Agentforce and Slack, with a 37-skill Salesforce plugin in pilot (Salesforce).
• OpenAI served Cursor a termination notice effective 12 November; Cursor said OpenAI was near 5% of its traffic and that Anthropic raised Claude's allocation the same day (Business Insider, Bloomberg Law).
• Nvidia has committed $26bn over five years to open-weight model development, alongside $13bn for Hugging Face and $7bn for Poolside's model factory (WIRED).
• Martin von Zweigbergk, creator of the Jujutsu version-control system, joined East River Source Control as chief technology officer (ERSC).
• Several vendors are now billing agents by result rather than by token — Intercom at $0.99 a resolution, Zendesk at roughly $1.20–$1.50 for LLM-verified resolutions within 72 hours, against Salesforce's original $2 per session regardless of outcome (The Next Web).

Platforms shifted

• Mozilla shipped an ad blocker for Firefox on iOS (Mozilla).
• The ChatGPT/Codex desktop app was found to bundle a complete copy of LibreOffice (Simon Willison).
• Google Play blocked Aurora Store, which GrapheneOS users rely on to install apps (AuroraOSS).
• Figma published Workflow Lab, its account of moving work between design and code with agents in the loop (Figma).

We cut a run of design-press pieces this morning — a logo backlash, a rebrand, an ad's behind-the-scenes video — because none of them changed anything you would do today. We also cut three tools from a "best of the week" list after noticing every one of them was owned by the person doing the recommending.