Meta just put its AI agent on a keychain

Top engineers at Anthropic and OpenAI say AI now writes 100% of their code.

If you're not using AI, you're spending 40 hours doing what they do in 4.

These 100+ Claude Code hacks fix that and help you ship 10x faster.

Sign up for The Code and get:

Claim your free playbook

1.  Meta turns Muse into hardware you carry. Meta announced Muse Charm at Connect: a pendant-sized device with a small screen showing your Muse, a fingerprint sensor on the edge and no phone required. Press it and the agent answers a question or runs an errand. Zuckerberg said only a few have been built and it ships for the holidays, with no price yet. Muse is also coming to every pair of Meta glasses. Tom's Guide has the full keynote; the 30-second demo shows the idea faster than words.

2.  The Quest gets replaced by a pair of glasses. Meta also unveiled Meta VR Glasses at $1,299 for spring 2027: about 100 grams, roughly a fifth of a Quest 3, with the computing moved into a separate puck, eye and hand tracking instead of controllers, and a keyboard it projects onto any flat surface. It plays the full Quest library and 3D Marvel films through Disney+. Camera-free Ray-Ban Meta audio glasses arrive October 13 at $349, per Techloy.

3.  ChatGPT Voice gets your inbox, calendar and Slack. OpenAI upgraded ChatGPT Voice to run on the GPT-6 family and connected it to email, calendar and Slack. You can now say out loud "draft a reply to Sarah and send it," "what is on my calendar Thursday," or "build me a spreadsheet of these numbers," and it does it, on your phone or desktop. It is live in ChatGPT Work now, and the demo is the best two minutes on what a hands-free workday looks like.

4.  Claude finds an enzyme system nobody had named. Anthropic reported that about 950 Claude agents, searching DNA databases for 21 hours, spotted a previously unknown enzyme system in bacteriophages with a repeating structure reminiscent of CRISPR. Human scientists then confirmed it in the lab and Feng Zhang of the Broad Institute called it intriguing. Nobody knows yet what it does. Anthropic says it is sharing early on purpose and that all the bench work was done by people.

The setup. Tulsee Doshi, who runs product for Gemini models at Google DeepMind, sat for a long interview this week, and the most useful part was not the Gemini 4 confirmation. It was her description of where the industry is: on a Pareto frontier, her phrase, where the question is no longer who has the smartest model but who delivers a given level of intelligence fastest and cheapest.

The evidence. Look at the week. Anthropic shipped Opus 5.5 at 40% lower cost. OpenAI shipped Sol and Luna at half the price of the last generation. Google positions Gemini 3.8 Flash as the sweet spot for teams watching token budgets. Three companies, three days, all competing on the same axis, and none of them leading with a benchmark.

The Google defense. Doshi pushed back on the idea that Google is slow, and the numbers back her: Gemini's last three models shipped within three weeks of each other, each incorporating developer feedback, in what Demis Hassabis calls relentless shipping. Whether or not Google wins, the cadence is real.

What it means for you. If price and speed are now the competition, the winner for most work is whichever model is good enough at the lowest cost, not the best one. That flips how you should pick. Route drafting, summarizing and classification to the cheap tier; save the frontier models for work where a wrong answer is expensive. Doshi's point is that a year ago the cheap tier was not good enough for that. Now it is.

The reframe. The barrier to entry is so low, Doshi said, that right now may be the most exciting time to build. Read that from the other side: if intelligence is a commodity priced per million tokens, the moat is not the model. It is what you connect it to, which is exactly why Meta spent its keynote on hardware and OpenAI spent its week wiring Voice into your inbox.

Burn the tokens. A Reddit user asked Claude to waste the rest of the week's usage allowance on something absurd, and Claude obliged at length. A month ago people were rationing these limits; now they are setting them on fire for fun.

The Wang unfiltered hour. Meta's chief AI officer Alexandr Wang was accused of using a social media team and responded with a posting spree unusual for an executive at a trillion-dollar company (1M+ views). The posts live on Superhuman's archive only, so no link.

Opus 5.5 in the wild. A roundup of eight projects built on Anthropic's new model, mostly animation and long-form writing, is circulating this week. If you have not tried it since the upgrade, the writing is the change to look for.

Mad Max, AI edition. An AI-generated 60-second Mad Max parody summarizing the whole industry, players, feuds and motives, passed 3M views. It is unlinkable outside the archive; you will see it anyway.

Three tools worth a look. Solid gives agents their own accounts, computers and budgets so they stop borrowing your credentials, ToneBird writes replies that remember which relationship they are for, and AgentScore scores your agent daily so you can tell whether an update helped.

Storyboard a Reel you can shoot tomorrow

A vertical video is easy to imagine and hard to plan, which is why most never get filmed. This prompt makes ChatGPT or Claude act as a director: it produces a shot-by-shot table with timestamps, camera moves, on-screen text and a voiceover script, plus a props list and a checklist, all sized to whatever gear and budget you have. The output is meant to be handed to whoever is holding the phone.

Swap [PRODUCT, BRAND OR TOPIC] with something specific like "our refrigerator water filter", "a Sunday farmers market" or "why I quit my job", and be honest in [AVAILABLE RESOURCES]; "one phone, one person, my kitchen" gets a far more usable board than a blank. Advanced tip: ask for two versions, one with a talking-head hook and one with a product-only hook, and shoot the first three seconds of both before committing.