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 · 15 STORIES

AI news, 10 November 2026

From the community

Gemini 4 Carbon is now on the paid Gemini API and rolling out to Google AI Ultra

Google's Gemini 4 Carbon, which Google calls its "strongest model", was announced on 28 October, but only for the Fairwind cybersecurity program. As of 10 November it is on the paid Gemini API (AI Studio, Android Studio) and Gemini for business, and is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity. It is not on the free API tier. Lower tiers have no date, and Google has published no country rollout schedule. Introductory API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens through 30 November 2026, then $4 and $20 from 1 December. It has a 1 million-token context window. Antigravity is previewing a feature where a team of agents splits up a project and checks each other's work. The Verge says the introductory API price is about a quarter of the price of Claude Fable 5.2, but cost per task can differ.

Why it mattersIf you already pay for the Gemini API or Google AI Ultra, this is the new top model, at a temporary introductory price that doubles on 1 December. It is not arriving on the free API, and there is no date for lower Gemini app tiers.

Sourcesblog.google ↗ · aistudio.google.com ↗ · theverge.com ↗ · x.com post ↗

Anthropic launches Claude Whiteboards, Claude Sheets and Claude Data

Anthropic has launched Claude Whiteboards and Claude Sheets in beta on all paid Claude plans (web, desktop and mobile). Whiteboards are shared boards that Claude organizes into timelines and deadlines, and that can be turned into a doc or a deck. Sheets build working spreadsheets with tabs, formulas, formatting and dropdowns, and Claude explains each edit. Claude Data (formerly Claude Dashboards) is now generally available on paid plans: it connects to a data warehouse or CRM and answers with dashboards, written data stories or narrated videos. Also launched: Claude Meetings, where Claude joins video calls as a visible participant (beta, Team and Enterprise only), and interactive connectors that open the full Adobe, Atlassian and HubSpot apps inside Claude. The free plan does not get Whiteboards, Sheets or Data.

Why it mattersIf you already pay for Claude, these are new work tools inside the product rather than a new model. Whiteboards and Sheets are still in beta, and Meetings is limited to Team and Enterprise. The free plan is unchanged for Whiteboards, Sheets and Data.

Sourcesclaude.com ↗ · techcrunch.com ↗ · theverge.com ↗ · x.com post ↗

Google Antigravity gets a layout for research and writing, not just code

Google's Antigravity agent app now has a Knowledge Work layout tuned for research, writing and producing documents. It keeps the same agent, plugins and MCP tool connections, but hides the terminal and git panels that coders use. You switch to it in Settings, under the General tab.

Why it mattersNon-developers on a team can use the same agent setup as engineers without a coding-style screen getting in the way, which makes Antigravity a more realistic option for marketing, ops and research staff.

Sourcesx.com post ↗

Hermes Cloud agents now run around the clock and open in any browser

Nous Research's Hermes Cloud agents now open in a browser tab with the same interface as the Hermes Desktop app. You pick a name and a model in Nous Portal, and Nous hosts the agent in its own container, running 24/7 so it can do scheduled jobs and reply on any channel while your laptop is closed. Nous says Plus, Super and Ultra plans each include one Cloud agent.

Why it mattersA small business can have a personal agent that keeps working overnight without leaving a computer switched on, and reach it from a phone. Plan details are Nous's own description.

Sourcesx.com post ↗

Meta's Muse Ember: a cheap 'worker' model for the jobs agents repeat all day

Meta launched Muse Ember, its smallest and cheapest model, built for the high-volume parts of agent work: tool calls, coding subtasks, routing, data extraction and reading long documents. It costs $0.10 per million input tokens and $0.50 per million output tokens, takes text and images, and handles up to a million tokens of context. Meta says it matches Claude Haiku 5.5's score at about a third of the cost per task. The Next Web's coverage notes it trails Haiku on one coding benchmark. It's also available on Google Cloud. Fact or hype: Partly fact. The launch and prices are confirmed; "matches Haiku at a third of the cost" is Meta's own claim, and independent coverage shows it behind Haiku on at least one coding test.

Why it mattersThis is how agent bills come down: a stronger model plans, and a cheap one handles the volume, such as sorting support tickets or pulling data out of invoices. Test it on your own repetitive task before relying on the comparison.

Sourcesx.com post ↗ · research.meta.ai ↗Partly fact Hype or Fact? check

Why your AI video looks fake, and an 8-step fix

ElevenLabs' creative-suite account, @ElevenCreative, published a guide on the six usual reasons AI video looks fake: lighting nobody specified, a generic setting, characters that are too perfect, vague physics, leaving sound to the video model, and skipping post-production. Its fix is an eight-step workflow: plan the details, generate the set, build a character sheet from a Claude-written prompt (it suggests saving that prompt as a Claude skill), add a voice, combine image, character and voice into a video, add sound effects, then colour-grade in an editor. It includes reusable prompt templates, including a timed, shot-by-shot video prompt with a checklist. Note: this is vendor content; the workflow is built around ElevenLabs' own Flows product, though the prompt templates work with other tools.

Why it mattersIf you make product or social videos with AI, these are the details that separate "obviously AI" from usable. Copy the character-sheet and shot-by-shot templates even if you don't use ElevenLabs.

Sourcesx.com post ↗ · elevenlabs.io ↗

OpenAI's launch-day news: dots on several computers, and a 'Control Room' for agent jobs

On the day of Anthropic's launch, @OpenAIDevs announced that you can now connect several of your own computers to the same ChatGPT "dot" (its always-on agent) and choose which one runs a task, so it can build and test on a second Mac while you keep working on the first. Separately, OpenAI said it has built "Control Room", one central system that tracks every research workload it runs, including agents generating data for training and evaluations, and can stop them quickly if a safety or security concern comes up. @kimmonismus is hoping the "Tomorrow." teaser from 8 November means something bigger is still coming; nothing has been announced.

Why it mattersGiving an agent its own computer is a practical way to keep it from getting in your way; use a spare or work-only machine. And Control Room is a pattern worth copying at small scale: one place where you can see, and switch off, every agent job you run.

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

Reports: Anthropic could start marketing its IPO this week

@kimmonismus passed on a Bloomberg report that Anthropic could start marketing its stock market listing as early as today (Tuesday), for a Nasdaq debut the following week, and that investors have discussed a valuation of $1.1 to $1.3 trillion. CNBC also reports the IPO could come as soon as next week. Anthropic hasn't announced a date.

Why it mattersNothing to do yet. A listed Anthropic will publish more about its finances and plans, which helps if you're weighing how much of your workflow to build on Claude.

Sourcesx.com post ↗ · cnbc.com ↗

Greg Isenberg's read: 'a night shift for every small business'

@gregisenberg turned the Fable launch into business predictions. Anthropic's own table shows Fable 5.2 completing 41.3% of tasks on the Remote Labor Index, a test built from real freelance jobs, against 21.3% for Opus 5.5. Greg argues agencies may move to pricing per deliverable, and that a model able to work alone for 10 to 12 hours is like a night shift for small businesses. These are his opinions built on Anthropic's tests, not results from real client work.

Why it mattersIf you sell hours (an agency, a freelancer, a consultancy), it's worth thinking now about which deliverables you could price as outcomes. Just treat the percentages as lab benchmarks, not proof of what it'll do on your jobs.

Sourcesx.com post ↗

Anthropic says Fable 5.2 built a prototype operating system with 1,000+ subagents

@kimmonismus highlighted an experiment on Anthropic's launch page: one Anthropic engineer asked Fable 5.2 to design a prototype operating system from an empty folder. Anthropic says the model coordinated more than 1,000 subagents working around the clock for eight days, and produced a system that boots on a laptop, with drivers for the disk, keyboard and network, file storage, windows, a login screen and an installer. This is Anthropic's own account of an internal experiment; nobody outside has tested the result, and it's a prototype, not a product.

Why it mattersThe lesson isn't "build an OS". It's that long jobs now get done by one model managing many smaller agents. Before you try anything similar, make sure you can see what each agent is doing and stop it.

Sourcesx.com post ↗ · anthropic.com ↗

Anthropic's free plugin that tidies up your Claude Code setup

@deanwperkins pointed to claude-code-setup, an official Anthropic plugin that looks at your project and recommends which automations to add (hooks, skills, MCP servers, subagents and slash commands), with steps for setting each one up. It only reads and recommends; it doesn't change anything. Install it with "/plugin install claude-code-setup@claude-plugins-official".

Why it mattersIf your team's Claude Code setup grew by trial and error, this is a quick, low-risk way to see what you're missing before you build more custom skills.

Sourcesx.com post ↗ · github.com ↗

earshot: a free tool that lets coding agents 'hear' sound

earshot is a free, open-source MCP server by Leonxlnx that runs on an ordinary CPU and gives agents like Claude Code a sense of hearing. It can find sound events in a file, check whether audio lines up with the cuts in a video, critique the timing and generate sound effects in code. It was shared by @shmidtqq, who isn't the builder. I haven't tested it.

Why it mattersSound landing slightly late on a cut is one of the giveaways in AI-made video. If you produce reels with an agent, this is a way to have it check the audio instead of you scrubbing through by ear.

Sourcesx.com post ↗ · github.com ↗

Drex Max: a 'decision model' that picks rather than writes

Nace AI's Drex Max is a 27-billion-parameter open model that chooses among options you give it (with text or images) instead of generating text, which suits classification and routing. @kimmonismus shared the company's claim of 183 milliseconds per answer against 3.5 seconds for Opus 5.5, and @dr_cintas called it "#1 on Artificial Analysis". That "#1" is on Artificial Analysis' Decision Index v2, a specialist ranking where it's tied first, not on the general model leaderboard. Artificial Analysis measured about 275 milliseconds per task, slower than the post says. Fact or hype: Partly fact. It's real and tied first on one specialist index; the speed comparison is the company's own, and an outside measurement was slower.

Why it mattersIf you sort large volumes of tickets, leads or documents into fixed categories, a small decision model could be much cheaper and faster than a chat model. It's niche, so test it on your own labels.

Sourcesx.com post ↗ · artificialanalysis.ai ↗Partly fact Hype or Fact? check

Brief the problem, not your solution, says @mattpocockuk

@mattpocockuk's tip for when you let an agent interview you before it builds: explain why you want something, not what to build. "Admins can't see an overview of the important data" works better than "build an analytics panel with graphs and filters", because if you lead with your solution, the agent sticks to it, and if you lead with the problem, it explores better options.

Why it mattersIt's a ten-second habit that improves any AI brief, coding or not, from a landing page to a report.

Sourcesx.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.