The AI Enthusiast.

Meta built the whole stack this week. The last piece was a way to get paid. #

Meta Connect 2026 · September 23–24, 2026

Connect was the Muse show. A Muse Charm pendant shipping in December, camera-free Ray-Ban Meta Audio glasses from $349, and Zuckerberg's clearest statement yet on how any of it makes money: a small transaction fee paid by merchants, not users, and not ads. With more than 1,500 developers applying to build Muse connectors in the first week, Meta is assembling the first large-scale app store for AI agents — on hardware and terms it controls, after a decade of chafing under Apple's.

The hardware nobody expected

Zuckerberg closed the keynote with the Muse Charm: a keychain-sized pendant with a 2-inch screen showing a "Jolly" avatar, a fingerprint sensor for instant voice activation, a small camera, at least three microphones, and — the specification that matters — a built-in 5G modem, so it does not pair to a phone at all. Only a handful of units exist. Meta is targeting December and has not announced a price. It follows Meta's December 2025 acquisition of the wearable-AI pendant maker Limitless.

Read that spec sheet again with the business model in mind. A device with its own cellular radio is a device that does not sit inside anyone else's operating system, does not appear in anyone else's app store, and cannot be governed by anyone else's review guidelines. The 5G modem is not a convenience feature. It is the point.

The money, finally said out loud

The fee comes from merchants, not users, and Meta is explicitly not putting ads inside the agent. To see how surprising that is, look at the base: Meta booked $59.36B of advertising revenue in Q2 2026 out of $60.80B total. For twenty years the company has been almost entirely an ad business, and it has just described a second one.

Chief AI Officer Alexandr Wang laid out the commerce surface: Muse takes PayPal and can shop at Walmart, Best Buy, Gap, Sephora, Wayfair and American Eagle, with Expedia for travel and Instacart for groceries. How large the fee is, and how it is calculated, were not disclosed — Meta has not explained the mechanics of collecting from merchants or what share it keeps. Amazon has also been reported to have blocked Muse from shopping on its platform; treat that as unconfirmed, but note that Amazon, Google and Apple all run their own commerce layers and none of them will sit still while Meta builds a toll booth in front of checkout.

Connectors are the new apps, and Meta's cut is the new 30%

Zuckerberg has paid Apple's 30% for years, and App Tracking Transparency kneecapped Meta's ad targeting in 2021. His position since the Rogan interview has been consistent: build the fundamental technology yourself so no platform provider can ever tell you "nah, you can't build that." Line the products up and the pattern is unmistakable — glasses against iPhone, Muse against Siri, Charm against Apple Watch. Meta is not competing with Apple's AI. It is competing with Apple's position.

Meanwhile, OpenAI is still building the object

OpenAI's Jony Ive device — born from the $6.5B io Products acquisition, a pocket-sized, screen-free, voice-first "third core device" — has been delayed to 2027, and the "io" name was dropped after a legal dispute. Meta's Charm ships this December. The difference in approach is the story: OpenAI is building an object; Meta is building an ecosystem the object plugs into. One of those is much harder to copy.

The camera comes off. The battery doubles, the backlash goes away, and the price drops a hundred dollars. #

Ray-Ban Meta Audio · pre-order now on ray-ban.com · prices verified September 25, 2026

Ray-Ban Meta Audio is Meta and EssilorLuxottica's first audio-only AI glasses: the same Ray-Ban frames, with the camera deleted. Ray-Ban's own product copy states the trade plainly — "By removing the camera, you get up to 12 hours of battery life for hands-free calling, personal listening, Meta AI and more." They start at $349, against $449 for the camera-equipped Gen 3 Wayfarer. Two styles launched: the Clubmaster (model RW7004, 6 colors) and the Burbank (5 colors).

What Ray-Ban is actually charging

The $349 figure is the entry price, and the lineup gets more expensive quickly once prescription and Transitions lenses enter the picture. Checked against ray-ban.com on September 25:

Why subtracting a feature was the right call

The spec that sells these is the one Ray-Ban leads with: up to 12 hours on a charge, roughly 48 with the case, at 43 grams, with six microphones and open-ear audio for calls and hands-free Muse. That number is only reachable because the camera is gone — image sensors and the silicon behind them are what drained the earlier models by mid-afternoon.

All-day battery is the difference between a gadget and an accessory. An always-on agent needs always-on ears, and a pair of glasses that dies at 2pm cannot be the interface to one. Dropping the camera also quietly retires the "perv glasses" problem that has followed every camera-bearing face computer since Google Glass: nobody in the room has to wonder whether they are being recorded, because there is nothing to record with. Meta gets the conversational AI, the wearer gets a device nobody flinches at, and the price falls a hundred dollars below the camera model.

The volume argument is already made. Meta and EssilorLuxottica sold more than 7 million smart glasses in 2025. This is not a concept product looking for a market; it is a product line segmenting one it already has.

Muse gives every user their own Linux computer. And exactly one conversation. #

Launched September 8, 2026 on iPhone, Android, web and WhatsApp · Mac app September 17

Muse is now the No. 1 free app on both the US App Store and Google Play — more than 902,000 downloads in six days, 4.9★ from roughly 35,000 ratings. The architecture is half the story: each user gets an isolated Linux VM holding the agent, a browser and their connected accounts, with a Sentinel watcher approving everything before it touches the internet. The interaction model is the other half, and it breaks from every chatbot before it — one persistent conversation, not a pile of threads.

The machine underneath

Muse Secure VM gives each user one dedicated, isolated Linux computer containing the agent, its own browser, and their connected accounts — Gmail, Google Workspace, Instagram, Facebook, Threads, Apple Health, Android SMS. A separate Sentinel process must approve anything before it reaches the open internet, and the agent never directly sees passwords or payment details. The stack runs on Muse Spark. Connect added live video avatars, a dedicated Muse email address, and access from the smart glasses.

The deal

100 million tokens per week, free, with paid tiers at $20/month for 500M and $100/month for 3B. Make it free enough to become a habit, then monetize the transactions rather than the tokens. The Mac app that followed on September 17 does not just read files — it can take over the Mac and operate any app on it, reaching into Files, Mail, Messages, Calendar and Notes with opt-in permissions, plus Box and GitHub, and it keeps working in the background after you close the window.

The engagement number is the one to watch rather than the download number. Sensor Tower estimated 560,000 daily active users after 11 days, and Meta stock rose 11% on Monday, September 21 as investors read the early adoption data, with Wells Fargo and KeyBanc both raising price targets. Downloads measure curiosity; daily actives measure habit, and habit is what a transaction fee is underwritten by.

One chat, not many threads

• No filing system. You do not start a new thread with a human assistant when the topic changes. You just keep talking.
• Memory survives the topic change. Nothing is stranded in a dead thread — a recipe reel becomes a grocery list becomes a dinner party, with the dietary restrictions still remembered.
• Proactive work needs a home. An agent that acts while you are away has to report somewhere. One conversation is an inbox; many threads is a filing cabinet nobody opens.
• It is the interface consumers already live in. WhatsApp and iMessage are one thread per person. Muse is literally inside WhatsApp.

Instinct raised $250M without publishing a single number. Then demonstrated that revoke is not delete. #

Series B announced August 26, 2026 · invite-only beta since February

Instinct — the invite-only personal agent you text or call, from Spear Street Technology — raised $250M at a $2.5B valuation in a Series B co-led by Index Ventures and Benchmark, up from roughly $100M weeks earlier. Early users report genuine capability alongside exactly the trust incidents Meta spent its architecture budget trying to design around.

The product and the money

Founder Noah Shinn, 23, previously at Sierra. The pitch is that it is "trained to use a phone and a computer… in the same way that humans do." There is no app to learn — you text or call it over iMessage, and it connects to email, messaging, screens, audio, location and devices. It remains invite-only; the company says it is scaling compute. The raise brings it to $350M total, and it arrived before any published revenue, user count, retention figure, price, or even disclosure of which model it runs on. One investor's summary: "OpenClaw for normal people."

The incidents

• Claire Vo disconnected Instinct from Google — and still received an AI email summary three hours later. 36 Gmail messages had been retained, and deletion required a manual request. Revoking access is not deleting data.
• An email was sent without explicit approval, and a researcher landed a successful email prompt-injection test.
• Terms have been reported to permit screen captures, keystrokes and passwords, and to allow training on user information unless the user opts out.

Instinct against Muse

Capability is converging. The divergence is who absorbs the trust risk — Meta's answer is architecture, Instinct's is velocity. Consumers will settle it with their inboxes.

Jev, one week later. Fastest adoption in AI Gateway history, and six clones by Thursday. #

TypeSafe AI · launched September 15, 2026 · follow-up to Issue 24

Issue 24 covered Jev's launch — "unstructured state in, typed probabilistic decisions out." What happened next is the story. Jev became the fastest-adopted model in Vercel AI Gateway history: roughly 13% of paid teams inside 24 hours, about 2x GPT-5.6's start and 6x Claude Fable 5.1's. And six open-source clones appeared within 48 hours.

Why a discriminative model is a different animal

A traditional LLM is a generative decoder: it predicts the next token, producing open-ended text. Powerful, but slow, expensive, poorly calibrated, and prone to mangling structured output. Jev is a discriminative classifier and ranker: it scores candidate options and returns typed probabilities. Fast, cheap, deterministic in schema, and genuinely calibrated — two years of stealth RLCD training aimed at making "80% confident" actually mean 80%. Different job entirely: deciding, not writing.

The clones

Six in 48 hours — Laya, DiffusionGemmaJev, Bespoke Nimble, SemIf/OpenJev, Jevlike and Kev-0.5B. Their shared hunch is that Jev's core is closer to an encoder with a ranking head than to a giant decoder. The architecture was commoditized in two days. Calibration, developer experience and distribution were not — which is the more durable lesson about where moats actually live.

Opus 5.5 and GPT-6 Sol and Luna landed the same day. Pricing is now a weapon, not a schedule. #

Launched September 22, 2026

September 22 delivered a same-day showdown: Anthropic's Claude Opus 5.5 against OpenAI's GPT-6 Sol and GPT-6 Luna, with aggressive price positioning from both sides. Two labs do not land flagship pricing on the same calendar day by coincidence.

• Claude Opus 5.5 and GPT-6 Sol/Luna launched the same day, both positioned on price.
• The launches bracketed a week of benchmark one-upmanship — see story 13.
• If your agent workload is high-volume and low-creativity, this is the week to re-run the arithmetic — and to look at whether a decision-layer model like Jev (story 5) should be taking those calls at all.

Claude agents reportedly found a novel enzyme family. So Anthropic built a wet lab to check. #

Anthropic · reported during the coverage window

Anthropic reported that Claude agents discovered a novel enzyme family — and then announced its own wet lab to run the experiments the agents design.

• An enzyme-family discovery is attributed to Claude agents. The figures rest on secondary reporting.
• The wet-lab launch moves agentic science from simulation into physical execution — the same direction as the Model Hardware Standard covered in Issue 22.
• Note the adjacency to story 11: a lab capable of running agent-designed biology is precisely the capability that makes a public weapons pledge load-bearing rather than decorative.

Nobody agrees what an agent is. Noma, Cowork and OctoAI are three different answers. #

Meta · Anthropic · reported during the coverage window

Three competing visions of the personal agent collided this fortnight — Meta's Noma, Anthropic's Cowork, and OctoAI — each answering the question "what is an agent" differently, and each pricing and permissioning itself accordingly.

• Assistant implies you approve each action, and the vendor's liability is limited. Coworker implies delegated authority and a review relationship. Infrastructure implies neither — it implies an SLA.
• The positioning battle across the three offerings intensified during the window, without any of them converging on a shared definition.

Google called it "Personal Intelligence" and attached a trillion dollars to it. #

Google · reported during the coverage window

Google framed its consumer AI push as "Personal Intelligence" and attached a figure large enough to reframe the whole category: $1 trillion in planned capital expenditure.

• "Personal Intelligence" positioning plus a $1T capex plan, announced during the window.
• Set against Meta's approach, the contrast is instructive: Meta is buying distribution — frames, pendants, WhatsApp. Google is buying capacity.

GPT-6 Astra routed a circuit board on video. A clean DRC is a milestone, not a manufacturing approval. #

OpenAI · launch page plus community demos

OpenAI's GPT-6 Astra posted 72.6% on OSWorld 2.0 against 65.7% for GPT-5.6 Sol, was reportedly classified "Restricted" under OpenAI's internal safety tiers, and then OpenAI published a condensed video of Astra doing PCB layout in KiCad — turning a schematic into a manufacturable board by placing components and routing copper.

• OSWorld 2.0: 72.6% for Astra against 65.7% for GPT-5.6 Sol — from OpenAI's launch page, offline set v2026.08.08, reported as a partial score.
• The "Restricted" classification is consistent with a model that can operate machinery and design hardware.
• The video is a 15-second condensed playback: "GPT-6 Astra performing printed circuit board (PCB) layout in KiCad, turning an electronic schematic into a manufacturable PCB by placing components and routing copper connections."
• Community runs went further. The standout is an independent F405 FPV flight-controller build from a single prompt through to a finished board.

Anthropic says it will never build weapons. Write the date down. #

Anthropic · reported during the coverage window

Anthropic made an explicit public commitment: it will never build weapons. Not a usage policy clause, not a tier in an acceptable-use table — a bright line stated in public.

• The value of a pledge like this is entirely in its specificity. "Weapons" covers a great deal of ground, and the interesting question is where the boundary sits on dual-use work and defense contracting.
• It follows the pacing argument from Issue 24, where Anthropic committed unilaterally to embedded third-party evaluators — the same pattern of acting first and asking for regulation second.

Plugin4Shell. The agent stack's soft underbelly is, as always, somebody else's code. #

Security · reported during the coverage window

Plugin4Shell — a new attack surface in the agent plugin supply chain — surfaced this fortnight, with a name chosen to make exactly the comparison you think it is making.

• Every capability added in this issue enlarges the target: a Muse connector, a Mac app reaching into Mail and Files, an agent with a merchant payment path.
• The defensive questions are the boring ones — where did this plugin come from, what can it reach, what is it allowed to send outbound, and who reviews the next version.

GPT-6 and Fable 6 traded benchmark claims. No independent evaluator had scored both. #

OpenAI · Anthropic · reported during the coverage window

GPT-6 and Claude Fable 6 traded benchmark blows across the window, with vendor-reported numbers doing all the talking and independent evaluation lagging behind as usual.

• Competing claims from both labs during the window; at time of writing no independent evaluator had scored both models on the same harness.
• The practical move is unchanged: build a small evaluation from your own workload and run both. It will disagree with both marketing pages, and it will be the only number that predicts your bill.