Anthropic's new top model scores highest and costs $26.54 a task. Its own launch post tells most people to start with the older one.

Tools

Build & ship

• ⭐ OpenRouter Auto Code — Set your model to openrouter/auto-code and it picks the model and reasoning effort per request, using live price, provider health and cache status. A general openrouter/auto router launched alongside it. · No extra charge · you pay the price of the model that serves the request
• ⭐ Ace — A GitHub Next experiment for building software with agents and teammates from your own Mac. Each piece of work gets its own git worktree and lands as a pull request. Agents run as you, with a shell, so it is not a sandbox. · Open source (MIT) · bring your own Anthropic or OpenAI key
• Exact — Expo's new way to write an app once and run it natively on web, macOS, iOS and Linux, with no webview, while an AI agent builds, runs and tests it on each. Pre-1.0 and changing daily. · Open source (MIT)
• Floot — A connector for Claude, ChatGPT or Cursor that builds a full app from chat, with backend, database, sign-in and hosting, and publishes to web and app stores. · Free · Pro $25/mo · Power $100/mo
• PI-Desktop — A local-first desktop workspace for coding agents on macOS, Windows and Linux. Plugins add panels, tools and MCP servers, and you can switch between cloud and local models. Early preview. · Open source (LGPL-3.0)
• CUI — A dependency-free C99 UI library with about 190 components that draws its own widgets on Windows, macOS, Linux and the web. A complete app is about 250 KB. · Open source (MIT)

AI & agents

• ⭐ Claude Fable 5.2 — Anthropic's new top model, available today on the Claude Platform, AWS, Google Cloud and Azure. A fast mode runs up to 2.5× faster. · $8 / $40 per million input / output tokens · fast mode $16 / $80
• ⭐ Muse Ember — Meta's smallest model, built for tool calls, routing, extraction and long documents. 1M-token context, with text, image, audio and video input. · $0.10 / $0.50 per million input / output tokens
• Stripe data for your agent — Connect Claude, Cursor or another agent to your Stripe data and ask why revenue moved. It shows the SQL behind every answer. · Queries API free until 31 Jan 2027, then per credit · Reports API free
• Perplexity Teammates — Always-on agents that each own an ongoing job, like research or bug triage, hand work to each other and message you in Slack when they need a decision. Web only for now. · Perplexity Max plan
• Tobira Site Agent — An agent on your website that answers visitors and visiting AI agents from your own materials, books meetings and sends you leads. · Free plan with your own AI key · Lite $49/mo · Start $169/mo
• Arete-27B — Nous Research's retune of Alibaba's Qwen3.8-27B. On Nous's own test it cut answers containing state propaganda from 22% to 1%, with coding scores roughly unchanged. Eval and training data are published. · Open weights (Apache-2.0)
• e-means — A drop-in k-means replacement from Meta FAIR that reaches lower loss about 10× faster. It ships inside Faiss. · Open source (MIT, as part of Faiss)
• Quest — An iOS assistant that builds an editable memory of your goals and relationships from your accounts, then talks you through decisions. It asks for a lot of access to your life. · Free (iOS)

Design & create

• ⭐ Edits for Mac — Instagram's video editor on the Mac, with a multi-track timeline, keyframes, colour wheels, LUT and font import, and 4K export with no watermark. Apple silicon only. · Free
• Wanaka — Describe a game and get a playable 3D version in the browser, then keep directing the AI or edit the world yourself and share a link. · Free · Pro $15/mo billed yearly
• Oh My PPT — A desktop app that turns a description of a talk, lesson or story into editable HTML slides, using your own models. Works offline. · Open source (Apache-2.0)

Growth & ops

• seam — Turns the Mac notch into a control strip for on-device dictation, music, a focus timer and your calendar, and shows when Claude Code, Codex or Cursor finish. · $19.90 once for 2 Macs · 48-hour free trial
• MeowDisplay — Use an iPhone, iPad or another Mac as a mirrored or extra display for your Mac, with touch control, over USB, Wi-Fi or Tailscale. · Open source (GPL-3.0)
• Someone Was Here — A menu-bar app that photographs anyone who uses your Mac while you are away and sends the photo to your iPhone. It recognises your face on-device. · Free · instant phone alerts $4.99 once

Resources

Steal the template

• The logo, headline, product-shot ad: Shlomo Genchin's one-format ad system, which had Claude research a client's audience and design 24 ads in one chat. The format is public; the full prompts are paid.
• WebMCP, explained with a to-do app: Ahmad Alfy adds seven agent tools to TodoMVC through HTML forms and JavaScript, with a live demo and code you can copy.
• 85% of AEO is won offsite: Maja Voje and Josh Grant's step-by-step playbook for getting cited by AI answer engines through creators and third-party sites. At Webflow, AI-search traffic converted at 24%, against 4% for non-brand Google search.
• A language tutor that refuses to speak English: Claire Vo walks through how William Imoh built native iOS and Android apps with Codex, with no Swift or Kotlin experience, and runs a company of about 25 people with a team of named agents.

Learn the craft

• The components of agent memory: Anthropic tests which memory designs help agents. Letting the agent reorganise its memories offline across 128 sessions, instead of 2, lifted its score from 13% to 48%.
• Turn agent failures into tests: NVIDIA's loop for mining agent traces, writing tests from the failures and checking each harness fix against them overnight.
• Securing AI agents at Uber: Uber Engineering on checking not just whether an agent may take an action, but whether the action matches what the user asked for.
• How Revolut cut its AI coding bill by 90%: David Abitbol's short conference talk on moving 2,000 engineers to self-hosted open models, and the budget rules that came with it.

Benchmarks you can re-run

• Co-Design Bench: 273 real app-editing tasks with expert-written hidden acceptance criteria. It scores both criteria passed and tasks fully done, and the gap between the two is the finding.
• BioBasicsBench: OpenAI's 100 routine wet-lab data tasks. The best setting, GPT-6 Astra at its highest effort, earned 66.7% of the credit.
• Project EVE: Anthropic put Claude into a replay of EVE Online's market and into team play with a hidden saboteur. The saboteur was almost never caught.
• SaaS sales efficiency at IPO: Sammy Abdullah's numbers for 158 SaaS IPOs, so you can compare your own new revenue per sales-and-marketing dollar.

Reads

☕ Under 5 minutes

• Unglamorous: Nikhil R lists the small, embarrassing work founders did early, from YouTube's $20 Craigslist ads to pump.fun's 3,000-plus cold DMs.
• How Grayscale made channel sales work: David Cummings on how Grayscale found the busiest hirers among its partners' customers and worked with the partner rep who owned each account.
• RIP Harvest: Kyle Kingsbury's time-tracking bill went from $13.32 to $133.98 a month. A reminder to check what your vendors can do to you.
• Outlook is rejecting mail without DMARC: since 5 November, small senders on custom domains are seeing hard bounces. Check your SPF, DKIM and DMARC records today.
• What does a positive AI future look like?: Patrick Collison asks what work and scarcity look like if AI goes well, and what happens if AIs have moral worth.
• AI transformation has three stages: Eric Seufert's sketch: add an LLM to every workflow, set efficiency benchmarks, then go on a token diet.
• On being an engineer in 2026: Shahar Glazner argues that agents raised the value of the engineers who direct them rather than replacing them.

🍵 5–10 minutes

• Fable 5.2 one-shotted my nephew's dream game: Lex Toumbourou, with early access, turned a child's drawings into a 12-stage game overnight from one prompt. The model wrote a bot to play-test every stage.
• Inside Clay's GTM machine: Clay's Osman Sheikhnureldin on the 17 workflows that score 250,000 accounts and give each rep a weekly top five. All figures are Clay's own.
• Locating a mysterious beep with Claude: Simon Willison had Claude Code build an app that turned five devices into listeners to find where a beep was coming from.
• OpenAI's internal research horizons: Basil Halperin fits a curve to OpenAI's own chart of research-task success and finds the time horizon doubling about every two months. The fit is his, not OpenAI's.
• The end of AI Frontiers: John Langford on why Microsoft Research's 50-person agent lab was folded into Microsoft AI, and why he left.
• 7 ways to handle your AI gross margin problem: Gokul Rajaram on who pays the token bill, with Intercom's $0.99 per resolution as the model.
• Why wouldn't the labs do it?: Josh Chapman argues the safest startups work where the labs won't go: physical operations, proprietary data and regulated work.
• Revenue only sort of matters to a strategic buyer: Jason Lemkin on why buyers pay for what your product becomes on their distribution, using Stripe's purchase of Privy.
• Formalizing CPython: David Andrews and Chao Zhang had Opus 5.5 agents write 613,318 lines of Lean that mirror CPython. Their bug list runs to 418 reports.
• State of a system: one button had 260 variants. Only 15 needed real thought, so the author argues for governing an LLM skill instead of a component kit.
• Why designers should welcome AI prototypes: Paul Boag says to treat a client's AI prototype as a working spec, not a finished design.
• What is malloc doing, anyway?: Ben Andrew traces 1,000 small allocations down to just two system calls.

📚 Longer, worth it

• AI is dramatically accelerating AI research right now: Ajeya Cotra on what METR's Anthropic report means and where the trend could lead.
• Data to deployment: Yoni Rechtman, after 60 calls across the data supply chain, on why data will split into cheap synthetic tokens and a scarce market for novelty.
• Rebuilding the global town square from scratch: X replaced three web servers with one. In a lab recording on a slow connection, the first post appeared at 2.7 seconds instead of after more than 22.
• The expectation economy: Jakob Nielsen's long case for designing like what users already know, with 37 guidelines at the end.
• Modern CSS might make your front-end more complex: moving logic from JavaScript into container queries and :has() can hide dependencies across the app.
• The rise of the one-person institution: Ryan Fedasiuk says knowledge workers should build systems, not documents.

Numbers

• $8 / $40: Claude Fable 5.2's price per million input and output tokens (Anthropic).
• $26.54: Fable 5.2's cost per task on the Artificial Analysis index, where it ranks first of 74 models with a score of 65 (Artificial Analysis).
• $0.10 / $0.50: Muse Ember's price per million input and output tokens (Meta).
• 74%: how much less per task OpenRouter says Auto Code costs than Opus 5.5 at a similar score (OpenRouter).
• ~2×: how fast METR estimates Anthropic is now improving its models, compared with March 2024 to October 2025, with a range of 1.3× to 3× (METR).
• 69.9% / 26.4%: criteria passed on an average task, and tasks fully completed, by the best model on Co-Design Bench (Co-Design Bench).
• 142 → 0: malformed tool calls in an NVIDIA agent example after its failures were turned into tests (NVIDIA).
• $0.66: median new annual revenue per sales-and-marketing dollar in the IPO year, across 158 SaaS IPOs (Blossom Street Ventures).
• $1.25T: OpenAI's reported valuation in a round that has raised $11 billion so far (Bloomberg).
• 16%: yield on CoreWeave's 2032 bond (CNBC).
• 46 of 47: decided data-center ballot measures this year where a majority favoured more limits (Ballotpedia).
• 179×: the total energy of answering in Pashto against English (3,147 kJ against 17.6 kJ), in an EMNLP 2026 paper (EMNLP 2026, "The Language–Energy Divide").
• 26.8%: rise in US business applications with no early hiring signs in AI-exposed industries, Q1 2024 to Q1 2026 (Liya Palagashvili, Mercatus Center).
• $0.35: median parking-app fee per session, across the 301 places in Park Graph's data that publish one (Park Graph).

What happened

Models & science

• Anthropic released Claude Fable 5.2 to everyone, and Claude Mythos 5.2, the same model without cyber and biology safeguards, to approved organisations (Anthropic).
• Meta released Muse Ember, its smallest model, on its Model API and Muse Code (Meta).
• Meta added Spaces to its Muse app: describe a gift tracker or trip planner and Muse builds it and keeps it updated (Meta).
• PSI says a system running an unreleased Anthropic model proved the Aaronson-Ambainis conjecture in quantum complexity, checked in Lean. It has not been peer-reviewed (PSI, The Quantum Insider).
• Nous Research released Arete-27B, a retune of Alibaba's Qwen3.8-27B, with its weights, data and evaluation (Nous Research).
• Meta FAIR released e-means, a faster k-means, inside Faiss (Meta FAIR).
• Google announced two new TPU systems, 8tx for confidential inference and 8i Max with two chips on one package. No availability date was given (Google).
• Infineon says it has made working power devices on 300 mm silicon carbide wafers for the first time, in its Villach fab. It is a process milestone, not volume production (Infineon).

Safety & policy

• METR estimated Anthropic's model progress has roughly doubled in speed. Most of the evidence it relied on is redacted (METR).
• AP reports the US government's AI safety testing center has no permanent leader, has stopped publishing research and is reporting informally to the White House, citing two former officials (AP).
• Nvidia's board opened an independent investigation involving CEO Jensen Huang and said it has no reason to believe company secrets were compromised (Nvidia).
• Google's Gemini help page says a subset of data from connected apps such as Drive and Gmail is reviewed by human reviewers, and asks users not to connect apps holding confidential information (Google).
• AI super PACs spent about $79 million across 68 House and Senate races this cycle (Tech Influence Watch).
• OpenAI committed $4.8 million for school meals in Effingham County, Georgia, and $6 million for childcare in Pike County, Ohio, near its data centers (OpenAI).
• Goldman Sachs limited Claude Code to its roughly 12,000 developers, taking it away from traders and bankers (Bloomberg).

Business moved

• OpenAI has reportedly raised $11 billion at a $1.25 trillion valuation, short of a $30 billion target (Bloomberg).
• Anthropic reportedly agreed to pay about $3.4 billion for 1.7 gigawatts of gas generators from OnePWR. Reuters could not verify it, and neither company commented (Financial Times, Reuters).
• CoreWeave's 2032 bond yield passed 16% and the cost of insuring its debt passed 1,289 basis points before easing (CNBC, Deutsche Bank).
• Apple lowered some engineers' daily AI spending cap from $300 to $200 (Bloomberg).
• The US added 165,000 jobs in October, about 70,000 of them in professional and business services (BLS).
• SemiAnalysis says the first 120 MW phase of China's Galaxy AI campus in Inner Mongolia was delivered within a year at about a fifth of US costs (SemiAnalysis).
• Microsoft folded its AI Frontiers research group into Microsoft AI in September. John Langford, one of its founding leaders, has left (John Langford).

Platforms shifted

• @GitHub in Slack and Teams is generally available: it can investigate bugs, implement changes and open pull requests from a chat thread (GitHub).
• GitHub began a public preview of a new pull request overview page with merge status and next steps at the top (GitHub).
• Stripe opened agent access to its analytics data for all users (Stripe).
• Perplexity launched Teammates, ongoing work agents, for Max subscribers on the web (Perplexity).
• Instagram released its Edits video editor for Mac, free (Instagram).
• The first macOS 27.3 developer beta puts Live Activities in the menu bar on Macs without a Dynamic Island (MacRumors).
• YouTube Playables now shows 24 accessibility tags for each game (YouTube).
• Payments Canada launched the Real-Time Rail, a 24/7 payment system (Payments Canada).
• Outlook.com has been rejecting mail without SPF, DKIM or DMARC records since 5 November, including from small senders (MailRoute).

We left out the week's viral app videos, a few Hacker News curiosities and a long run of sponsor placements.