A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 20 STORIES

AI news, 11 November 2026

From the community

TypeSafe's Jev came first. OpenAI and Google have followed with decision APIs

Google launched the Gemini Decisions API on 11 November 2026, powered by Gemini Decision 1. It follows TypeSafe AI's Jev, which opened in early access on 15 September 2026, and OpenAI's Decisions API, which the official changelog released in beta on 6 October and made generally available on 22 October. A decision model does not write a reply in prose. You give it some input and a set of bounded questions — yes/no, a choice from a list, or a score against a rubric — and it returns probabilities and a typed answer that software can use. That matters for agents and business automation such as triage, routing and policy checks, where you need a fast, machine-readable decision rather than a paragraph.

How they compare, from official pages checked on 11 November 2026. Latency figures are vendor-reported.

TypeSafe AI (Jev). Early access from 15 September 2026; TypeSafe's current docs list Jev 1.14, released 14 October, as the latest official version. Question types: noul, choice and score, plus box on Jev 1.14 (to locate something in an image). Inputs: text and PNG or JPEG images; it does not natively accept audio or video. Default context is 64k tokens per request and 32k per question, or 262k with an optional long-context header. There is no separate question cap; a request can include as many questions as those context limits allow. Price: $0.042 per 1 million input tokens, with output free. TypeSafe says end-to-end replies take 70–500 ms. The official endpoint is POST api.typesafe.ai.

OpenAI Developers (OpenAI Platform & Codex). Decisions API, built with GPT-6 Luna. Official changelog: beta on 6 October 2026, generally available on 22 October. Question types: predicate, choice and score, with up to 400 questions per request (200 in the public beta). Inputs: text, images and native JSON; official docs say files and audio are not supported. OpenAI does not publish a Decisions-specific context-window figure in the guide we checked. Price: $0.10 per 1 million uncached input tokens, $0.01 per million for cached input, with no output-token charge. OpenAI says it is up to 10 times faster than GPT-6 Luna through the Responses API, without giving its own millisecond figure.

Google AI Studio. Gemini Decisions API, powered by Gemini Decision 1, launched 11 November 2026 and available through the Gemini API in Google AI Studio. Question types: noul (yes/no; noul is TypeSafe's term), choice and score, up to 1,000 questions. Inputs: text, images, audio, video and PDFs, with 256K context. Price: $0.10 per 1 million input tokens and $0 for output, with a no-cost API tier in AI Studio for prototyping. Google says end-to-end latency is 150–250 ms. Support is coming soon to Gemini in Google Cloud.

Who should try which: if you want the lowest published list price, or you already use Jev, start with TypeSafe. If you already build on the OpenAI API and want a generally available endpoint with JSON input, prompt caching and US or EU residency, try OpenAI. If you need audio, video or PDFs in the same pass, a 1,000-question limit, or a free prototyping tier, Google's API is the one that documents those. We have not independently measured accuracy or latency.

Why it mattersAgents and automations spend a lot of time on small, bounded judgements. A decision API is meant to make those cheaper and faster than asking a chat model to write an answer. Google's launch means the three main commercial options are now TypeSafe, OpenAI and Google; pick on input types, limits and price, and test on your own examples.

Sourcestypesafe.ai ↗ · docs.typesafe.ai ↗ · docs.typesafe.ai ↗ · openai.com ↗ · developers.openai.com ↗ · developers.openai.com ↗ · blog.google ↗

Grok Bot 1.0: shared teammates in Slack, Microsoft Teams and meetings

SpaceXSI has launched Grok Bot 1.0, a major update that lets you share Bots and work with them in Slack, Microsoft Teams and video calls. The official Grok Bot account posted the 1.0 announcement on 11 November 2026, with this wording: "Introducing Grok Bot 1.0. AI teammates that work alongside your whole team. Share Bots, collaborate with them in Slack and Teams, invite them to meetings, and watch them learn and become more helpful." The post includes a 79-second demo video. The product blog, Introducing Grok Bot 1.0, is dated 2 November 2026. The official changelog lists version 1.0.1 on 11 November and says that entry also includes 1.0.0.

The blog describes three changes since the Grok Bot beta: you can share Bots and bring them into Slack, Microsoft Teams and video calls, and they learn from the people they work with; Bots can notice work that needs doing and offer to take it on; and they are described as faster and more reliable in the browser, and at writing documents and building presentations. Team Bots can be shared with specific people or groups. Official docs and the 11 November changelog say a Bot on a Team, Enterprise or Ultra plan can join a Google Meet, Zoom or Microsoft Teams call from a link, listen, take notes, answer in the meeting chat, and send a summary and action items afterwards. You can also ask the Bot for a transcript of a past meeting. Those docs do not say the Bot speaks aloud in the call. On Enterprise plans, an admin turns Meetings on or off in settings, so meetings are not on for every account.

Grok Bot is included in The X Subscription on Premium ($30 a month), Super ($100) and Ultra ($200), and is not included in Plus or Free, according to the official pricing page (checked 11 November 2026). Team seats are listed at the same prices. The blog says Team Bots are available on Team and Enterprise plans, and Meetings on Team, Enterprise and Ultra.

The launch video is a demo of a fictional headphones launch, not a record of a real customer project. It opens with "Introducing Grok Bot beta", then changes "beta" to "1.0". In the film, a Bot starts a channel on its own initiative, a user asks a Bot to join a 10 o'clock call and admits it to a video call that looks like Google Meet (no logo is shown), and "Research Bot" appears only as a silent participant tile: it does not speak, post in chat or show notes, and there is no recording indicator. On-screen text covers learning, recommendations and Bots working with other Bots, including a warning about a confidential slide before a deck is shared. The video does not mention pricing, plans, regions, dates or rollout, and it does not demonstrate a Bot speaking in a meeting.

Earlier coverage on this site: What's new in Grok Bot: shopping cards, order tracking, payments, custom modes and Team Bot logins and Grok Bot adds shopping, price comparison and order tracking.

Disclosure: The Independent AI uses Grok Bot to help run the site.

Why it mattersIf you already use Grok Bot, 1.0 is the shared-teammate release: one Bot can work with a team in Slack or Microsoft Teams, and, on the plans that include it, join a video call to listen, take notes and answer in the chat. Check your plan and, on Enterprise, whether an admin has turned Meetings on. The speed and reliability claims are the company's. Disclosure: The Independent AI uses Grok Bot to help run the site.

Sourcesx.com post ↗ · x.ai ↗ · x.ai ↗ · docs.x.ai ↗ · docs.x.ai ↗ · x.ai ↗ · docs.x.ai ↗

Movement Strategy built Chassis, an in-house AI analyst, and trained every team first

US social-first agency Movement Strategy has built Chassis, a proprietary AI data and reporting system that pulls owned social, paid and trend data. Campaign reported the launch on 11 November 2026. The agency is independent and not owned by a holding group, with about 185 people and hubs in New York, Denver and Los Angeles, according to CEO Jason Mitchell on The Ad Podcast on 28 August 2026. Staff can query Chassis — for example, what performed well last week and why. Mitchell also said on the podcast that AI agents browse social platforms daily, flagging algorithm changes and trends, and that there are "probably" hundreds of those agents. The system was built in-house; strategy director Emma Bunsness credits Taylor McDowell, the agency's director of business intelligence. Sung Kim, social media marketing and content strategy lead, told Campaign that all teams had been onboarded and were already using it "before we ever made a press announcement about it." Some clients have dashboard access: account director Daniel Gertis told Campaign that client-side access has been a selling point. Chassis is an internal tool with some client access, not a product readers can buy or sign up for, and it is not sold as standalone software. The time-saving figures are the agency's own claims and have not been independently verified. Mitchell told Campaign: "All that desk research and data analysis that used to take weeks can now be done in a couple of hours." Bunsness said a weekly report that once took two hours now takes 30 minutes. Mitchell also said humans remain "the tastemakers" who look at the output and decide whether they agree. Campaign did not publish adoption numbers.

Why it mattersThe useful lesson for other agencies and small businesses is not a product to buy, but a way of working: build on your own data, train team by team before you announce anything, and keep humans judging the output. Chassis is Movement's internal system; the speed-up numbers are the agency's claims, not independently verified.

Sourcescampaignlive.com ↗ · podcasts.apple.com ↗ · movementstrategy.com ↗

Buzz 1.0 is out of beta

Block's team workspace Buzz has its first stable desktop release: Buzz 1.0.0, published on GitHub on 10 November 2026 (not a pre-release). The builds cover Apple Silicon and Intel macOS (signed and notarized), Windows x64 (an unsigned installer) and Linux x64 (Ubuntu 24.04 deb and AppImage). Installed copies that use the updater get stable-feed updates once a release is promoted to it; a newer v0.7.0-preview build also exists on the preview channel. On 11 November, Jack Dorsey (@jack) wrote: "buzz 1.0 is finally here! the most flexible team workspace out there. for humans and agents, model agnostic, local or cloud, and infinitely extendable." "Local or cloud" is his wording. Buzz's official pricing page now lists a free self-hosted Community plan and a pay-as-you-go hosted plan that includes remote agents or managed cloud compute (paid remote agents and inference are available in the United States only); the 1.0 desktop release notes still describe agents running on your own computer while the app is open. It remains free and open source (Apache-2.0) if you self-host. There is still no mobile app. Enterprise is a custom agreement, with no published dollar prices. We have not tested 1.0 or the hosted plan.

Why it mattersIf you were waiting for a non-beta desktop build, this is it, and it now includes Intel Macs. If you want hosted agents, check Buzz's pricing page rather than relying on the "local or cloud" line in Jack's post: the official pay-as-you-go option is described there, and paid remote agents are US-only.

Sourcesgithub.com ↗ · x.com post ↗ · buzz.xyz ↗ · buzz.xyz ↗

Google DeepMind's AI proves Signal's encryption is secure down to the code

Google DeepMind says its AlphaProof Code framework has mathematically proven that Signal's messaging protocol is secure at the level of the actual code, a first for AI. Gemini agents wrote over 6 million lines of proofs in the Lean proof language, so human experts only had to review the critical 1%. These figures are Google's own.

Why it mattersProving software correct has been slow, expert-only work. If AI agents can do most of it, businesses could get far stronger guarantees for payment, messaging and infrastructure code. For now it is an early research result announced by Google.

Sourcesx.com post ↗

Google Antigravity's /swarm runs a team of agents from one prompt

Google added a /swarm command to Antigravity, its agent coding app, on desktop and in the command line. One prompt is split across specialised agents that plan, build and check the work together on a shared live message board, led by a coordinator agent. You can pin a different model to each role and follow progress on board and timeline tabs. Google showed it with Pelican Quest, an open-world 3D browser game of its Mountain View campus built with Gemini 4 Carbon and /swarm.

Why it mattersMulti-agent setups used to need custom frameworks. As a single command, it becomes easier to try splitting a bigger job, such as a prototype, a site or a research report, between several agents. More agents also means more usage, so watch the bill.

Sourcesx.com post ↗

Your ChatGPT Business plan now works inside Notion, Warp and other tools

OpenAI says companies on a ChatGPT Business plan can now use that plan inside third-party tools including Notion, v0, Amp, Kilo Code, Warp, Hermes Agent, OpenClaw and OpenCode, by connecting with Sign in with ChatGPT. Workspace admins decide which tools get access and can set usage limits, including for open-source tools.

Why it mattersIf your team already pays for ChatGPT Business, you may not need a separate AI subscription for every coding or work app. Admins keep control over which tools draw on the plan and how much.

Sourcesx.com post ↗

Google DeepMind's Gemini Robotics 2.5 learns hundreds of everyday tasks

Google DeepMind released Gemini Robotics 2.5, its latest model for physical robots. Google says it can learn hundreds of everyday skills, from folding laundry and pipetting liquids to serving food with tongs, and was trained partly in virtual practice spaces that Gemini 4 Argon builds from a single camera view. It also adds safety checks to spot hazards and steer away from people. In a proof of concept with Apptronik and Mercedes-Benz, the Apollo 2 robot picked and organised car parts for assembly with 96% accuracy, according to Google.

Why it mattersThis is still mostly lab and pilot work, but the factory test shows where physical AI is likely to land first: repetitive picking and sorting jobs on warehouse floors and assembly lines.

Sourcesx.com post ↗

MiniMax H3-Image: an open-weight model for 4K images in about 10 seconds

MiniMax released H3-Image, an image generation and editing model with open weights on Hugging Face. MiniMax says it makes 4K (16-megapixel) images in about 10 seconds, follows complex prompts closely and takes up to 16 reference images; these are the company's own claims. It is available through the MiniMax API, MiniMax Code and Hailuo AI, with 50% off 16MP+ images during launch week.

Why it mattersGood open image models are rare. For ad creative, product shots and e-commerce catalogues, open weights mean you can run it yourself or pick a cheaper host instead of depending on one vendor.

Sourcesx.com post ↗

Alibaba previews Qwen4-Max, its next flagship model

Alibaba's Qwen team says Qwen4 is coming and has opened a preview of Qwen4-Max on Alibaba's Token Plan, with early access in QwenWork, Qoder, the Qwen app and QwenStudio. Qwen pitches it at complex coding, knowledge work, front-end builds and professional software workflows, with stronger agent skills. The announcement gave no benchmark figures.

Why it mattersQwen models are among the most widely used lower-cost alternatives to US models. A new flagship is worth watching if you run coding agents on a budget, but this is a preview, so wait for independent tests before switching.

Sourcesx.com post ↗

OpenAI lets businesses claim their profile inside ChatGPT

OpenAI launched Businesses in ChatGPT in beta on 10 November. Through a new Business Hub, any eligible business where ChatGPT is available can claim its listing, verify by business phone, business email, a DNS record on its domain or Shopify, keep its details and product feed up to date, and see 28 days of insights: how often ChatGPT mentions or cites its website, website clicks, and "discovery topics" (the themes of answers it appears in). Eligible US ecommerce and online service businesses can also build a Business Agent that answers customers' questions in a separate, labelled chat, and can sponsor it as the destination for ChatGPT Ads. OpenAI says 1.2 billion people use ChatGPT each week (its own figure). @gregisenberg called it "the Google My Business moment for AI" and pointed out the gaps: you can't see the actual prompts, your competitors, or which chats turn into sales.

Why it mattersClaiming your profile is free and doesn't require ads or an agent. OpenAI says it won't make ChatGPT recommend you more often, but it does let you correct your details and, for the first time, see how often ChatGPT sends people your way.

Sourcesx.com post ↗ · openai.com ↗ · x.com post ↗

Anthropic launches its IPO roadshow: 825 million shares at $70 to $78

Anthropic officially started marketing its stock market listing on 10 November. It's offering 825 million shares at an expected $70 to $78 each, mostly new shares from Anthropic plus some from existing shareholders, and has applied to list on Nasdaq under the ticker "ANT". @kimmonismus worked out that this implies a fully diluted valuation of roughly $1.24 to $1.38 trillion, below the $1.5 trillion some had expected; that's his estimate, not Anthropic's. A listing date hasn't been announced by Anthropic.

Why it mattersNothing to do today. Once listed, Anthropic will have to publish much more about its finances and plans, which helps if you're deciding how much of your work to build on Claude.

Sourcesx.com post ↗ · anthropic.com ↗

Matt Pocock's 'Home Repo': one place for all your agent skills

@mattpocockuk's answer to working with agents across several projects: keep one "home" folder (repository) that holds all your skills, steering files and notes, always start your agent there, and list the other projects in its AGENTS.md file with where to find them. He says this gives you one place for all your agent setup, one place to run every agent from, and an easy way to make the same change across many projects at once. The post was bookmarked more than 1,300 times.

Why it mattersIf your skills and instructions are scattered across projects, each agent only sees part of how you work. One home folder is a free, ten-minute fix that also works in Claude Code, Codex or Cursor.

Sourcesx.com post ↗Fact Hype or Fact? checkSteps to try, cost and caveat ↗

Claude can now open Adobe, Atlassian, Box and more right beside the chat

Anthropic upgraded Claude's interactive connectors on 10 November: where a connector supports it, you can open the full tool beside your conversation or full screen, signed into your own account, and ask Claude to make changes you approve, which appear live. Examples include editing photos and PDFs in Adobe, writing a Confluence page in Atlassian, revising a Box file and comparing versions, and working in Notion and HubSpot. Anthropic says: "Interactive connectors run on MCP Apps, the open extension to the Model Context Protocol that lets an MCP server bring its own interface into AI applications." That is not the same as WebMCP, a separate W3C Community Group draft for exposing a website's tools to an agent. The side panel, full-site views, and opening or pinning a connector from Claude's navigation are rolling out on the web and in Claude Desktop to accounts with the new Claude experience; they are not on mobile yet. The 10 November article's starting tools include Asana, Canva, Carta, Dropbox, monday.com, Smartsheet, Vercel and Webflow, plus Adobe, Atlassian, Box, HubSpot and Notion. Figma, Harvey and Shopify are building integrations. In the Connectors Directory, connectors with this capability are marked with an "Interactive" badge.

What are MCP Apps?

MCP Apps is an official, optional extension to the Model Context Protocol (identifier io.modelcontextprotocol/ui). It began as SEP-1865, which is now Final; the stable specification is dated 26 January 2026. A server can ship its own interface as a ui:// resource, which a supporting AI application renders in a sandboxed iframe and talks to over JSON-RPC. If the host does not support the extension, the same tool is meant to fall back to ordinary text.

That is different from WebMCP. WebMCP is a separate W3C Community Group draft, not a W3C Standard, that lets a website expose tools to an agent through document.modelContext. Chrome has an origin trial from Chrome 149. With MCP Apps, the tool is a guest inside the AI app; with WebMCP, the agent is a guest on the website.

The official MCP Apps overview lists these supporting hosts: Claude, Claude Desktop, VS Code GitHub Copilot, Microsoft 365 Copilot, Goose, Postman, MCPJam and Archestra.AI.

For a small business or a non-developer, the practical change is less tab-switching: if a tool you already use offers an interactive connector, you can see its interface in the chat and approve changes there. Because this is an open extension, a vendor can build one interface meant to run in any host that supports it, not only in Claude. You still need to check that your account has the new Claude experience, and that the host you use actually offers the views you need.

Developers can start with the @modelcontextprotocol/ext-apps package and Anthropic's walkthrough, Get started with MCP Apps.

Limits: in Claude, the side panel, full-site views and pinning are rolling out only on the web and Claude Desktop, to accounts with the new Claude experience, and are not on mobile yet. For a directory connector, entry points and embedded sites need a Verified listing; a Community install does not get them. App tools and embedded sites in Claude are in public preview for developers, and purchases through in-chat views are not supported. The support article's "current interactive connectors" list also differs from the 10 November starting set above, so treat the article as the launch announcement and the support page as a directory snapshot that can change.

Why it mattersYou can stay in one chat and work in a tool you already use, instead of copying between tabs. An MCP App is not limited to Claude: Anthropic says it can run in any AI application that supports the extension. You will only see the side panel, full-site views and pinning once your account has the new Claude experience, and those views are not on mobile yet. Keep approvals on. Purchases through a connector's in-chat views are not supported.

Sourcesx.com post ↗ · claude.com ↗ · x.com post ↗ · support.claude.com ↗ · modelcontextprotocol.io ↗ · webmachinelearning.github.io ↗ · developer.chrome.com ↗ · claude.com ↗ · github.com ↗ · modelcontextprotocol.io ↗

Inkbox now runs your agents, not just their inboxes

Inkbox, which already gives AI agents their own email address, phone number, iMessage and WhatsApp, now hosts the agent too. Each agent gets its own cloud computer, wakes up when a message it handles arrives, sleeps when idle (keeping its files and scheduled jobs), and can run tasks on a schedule. You choose what wakes it and who it may talk to. The company also announced a $5.3 million seed round led by Peak XV. Shared by @dr_cintas.

Why it mattersThe hard part of an email or text agent is being there when the reply comes in hours later. This handles that without you running a server, which suits follow-ups, booking confirmations and lead replies.

Sourcesx.com post ↗ · inkbox.ai ↗

OpenBot 1.0: an open-source AI coworker workspace you can host yourself

CopilotKit released OpenBot 1.0, a workspace for AI coworkers that you can run on your own computer or servers. Each coworker gets a browser, files and tools; you choose the AI model; you can plug in agents your team already built with other frameworks; and company sign-in, permissions, approvals and an audit trail apply across the workspace. The app template is free and open source (MIT licence), but it relies on CopilotKit's separately licensed "Intelligence" component: free for personal and local use, paid for production. There's a desktop app for Mac and Windows. Announced by @ataiiam.

Why it mattersIf you like the idea of always-on AI coworkers but need your data on your own infrastructure, this is a credible option to evaluate. Check the licence before planning a company-wide rollout.

Sourcesx.com post ↗ · copilotkit.ai ↗ · github.com ↗

How Runway's marketing team uses OpenAI's image model

In a polished OpenAI video, Anthony Maggio (shown as Runway's Chief Product Officer) demos asking a chat to make eight bold still ads using a stored brand kit, then turning generated clips into a 30-second montage with music. He says the team went from about 70 to more than 900 ads a week; that's his claim and isn't shown. The post also mentions Runway's plugin for Codex and ChatGPT, which launched on 2 November and is for building apps with Runway's API; the video doesn't show it. I watched it: it's a testimonial, not a tutorial.

Why it mattersThe repeatable idea is the brand kit: store your logo, colours and fonts once so every AI-made ad starts on-brand, then ask for many variations and pick the best.

Sourcesx.com post ↗ · runway.com ↗

Agents make Goodhart's Law worse, says @mattpocockuk

@mattpocockuk warns that agents are very good at gaming whatever number you measure: when a measure becomes a target, it stops being a good measure. Any team that tracks something like the number of pull requests per person will get lots of pull requests, not more useful work.

Why it mattersIf you start measuring AI productivity, measure outcomes customers notice (things shipped, tickets solved, revenue), not activity counts an agent can inflate.

Sourcesx.com post ↗

A 'decision ticket' board to agree before agents build, from @mattpocockuk

@mattpocockuk is trying a shared board of "decision tickets" where a team talks through decisions before an agent starts building, instead of discovering disagreements at review time.

Why it mattersWhen agents build fast, the expensive mistakes are the decisions nobody agreed on. A short written decision list before you brief an agent saves rework.

Sourcesx.com post ↗

A well-known forecasting group now expects 'AI researcher' level AI in 2027

@kimmonismus flagged that the AI Futures Project, led by Daniel Kokotajlo, updated its forecast: absent a major slowdown, it now thinks AI able to match top human AI researchers will likely arrive in 2027, about five months ahead of its "AI 2027" scenario, citing METR's estimate that Anthropic's progress has roughly doubled in speed. This is a forecast, not a fact, and Kokotajlo himself says he is "scared, sad, and angry" about it.

Why it mattersYou don't need to plan around a date. The practical takeaway is that capabilities are arriving faster than most plans assume, so review your AI plans every quarter, not every year.

Sourcesx.com post ↗ · x.com post ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.