The big model news
Gemini 4 Carbon is out. Google is calling it its strongest model yet, built for the stuff that pushes a model's limits rather than everyday coding: multi-hour debugging, cybersecurity work, and open-ended scientific research. Its sibling Gemini 4 Argon, released last month, stays the default pick for regular software engineering. Carbon ships at the same introductory price as Argon ($2 per million input tokens, $10 per million output) through the end of November, and lands everywhere at once: the Gemini API, Google AI Studio, Gemini for Business, and Google AI Ultra across the Gemini app, Antigravity, and AI Mode in Search. Google is also previewing a /swarm feature in Antigravity that turns one prompt into a coordinated team of coding and verifier agents.
Anthropic matched the week's intensity from a different angle. Claude Meetings puts Claude into a live video call as a visible participant — it listens, drafts the plan being discussed, maps a timeline on a whiteboard, and can pull last year's numbers from a connected app while people are still talking. It launches in beta on Google Meet for Team and Enterprise plans. Alongside it: Claude Whiteboards (a shared visual workspace with timelines), Claude Sheets (ask for a spreadsheet, get a working one), and Claude Data, now generally available, which turns a connected data warehouse or CRM into dashboards, written reports, or narrated videos on request.
Adobe used its MAX event to push agents into its creative suite rather than build a chat assistant alongside it. Firefly gets Builder, a Figma-style prototyping tool in private beta, plus on-brand chart generation and object-level image editing. Adobe also shipped an MCP server so ChatGPT, Claude, or Gemini can call Photoshop and Premiere features directly, and new relighting and color-grading tools land across Photoshop, Premiere, and Lightroom.
OpenAI, meanwhile, pointed ChatGPT at commerce. Businesses in ChatGPT opens in beta: companies claim a verified profile through a new Business Hub, keep pricing and availability current, and see aggregated insight into how often they're mentioned. Eligible U.S. ecommerce and service businesses can also stand up a Business Agent that customers chat with directly inside ChatGPT — clearly labeled, and separate from ChatGPT's own answers.
Smaller but worth flagging: MiniMax open-sourced H3-Image, a 33.1B-parameter model that renders native 16-megapixel 4K images in about 10 seconds, free on Hugging Face. And Microsoft AI priced MAI Transcribe 2 at $0.10 per audio hour across 60 languages with built-in speaker diarization — a promotional rate running through the end of 2026.
Product Hunt
The theme this week was agents that come with their own body. Glasshand rents your AI agent a real iPhone or Mac sitting in a data center — connect Claude Code or Codex and it posts to TikTok, books an Uber, or builds and tests an iOS app the way you would, then gets wiped when you're done. It finished #1 on November 5th. A similar idea, NoMac, rents a cloud Mac to an agent at $0.90 an hour, billed per minute.
Wanaka Studio led the month overall — an AI game engine where an agent builds an AAA-style 3D game that you then edit in a synced visual editor, no code required, claiming roughly 80% fewer tokens than running Claude Opus plus Unity by hand. Product Hunt MCP topped the board on November 9th, putting Product Hunt's own community inside whatever assistant you already use. And Tobira Site Agent gives a website its own agent that negotiates with other companies' agents on your behalf — a small, strange preview of agents selling to agents.
Two launches stood out for being immediately useful without an upsell: Fleeep, a free notes app that pulls screenshots, bookmarks, and voice notes into one calm place and explicitly doesn't train on your data, topped the day on November 7th; and Instagram's free Edits for Mac brought 4K, watermark-free desktop video editing with an MCP hook that lets an agent assemble a rough cut from your Downloads folder.
On the stranger end, Agent Clan renders your Claude Code and Codex agents as a 3D team you direct with hand gestures — make a fist to summon them, pinch to pick one, pull right to approve a permission request. Free on Mac and Windows, mouse fallback included for when gesture control isn't practical.
GitHub Trending
Open-source agent memory kept climbing the charts. beacon.sh turns sessions across Claude Code, Cursor, Codex, OpenCode, and twenty-plus other harnesses into a shared, compounding knowledge layer — every new agent session starts smarter because of what previous ones logged. It was the top-scoring Product Hunt launch of the month, which tells you how much appetite there is for agents that don't forget everything between sessions.
Google DeepMind's AlphaProof Code is a different kind of trending repo: formal verification artifacts and a technical paper documenting an end-to-end proof that the Signal cryptographic protocol holds down to compiled machine code — six million lines of Lean proofs generated in four weeks. It's a narrow use case, but it's the clearest signal yet that AI-assisted formal verification is moving from research demo to something you can point at a real, widely-used protocol.
Worth your time
If you're choosing one model update to actually test this week, it's Gemini 4 Carbon for anything that currently makes you reach for a second pass or a longer context window — Google is explicit that it costs more compute to run and is built for exactly that kind of task, not routine coding. If you're choosing one Product Hunt launch to try, Fleeep is the lowest-risk pick: free, no training on your notes, and solves a real problem instead of wrapping a chatbot around one.