AI & design briefing for Friday, September 25, 2026

tldr

• Meta put Muse in hardware: Muse Charm, a keychain pendant about the size of an AirPods case with a 2-inch OLED for the avatar, 5G, two cameras and a fingerprint sensor, shipping in December, alongside camera-free Ray-Ban Meta Audio glasses at $349 on October 13.
• Anthropic opened Claude Marketplace with more than 2,000 connectors and plugins at launch, agents from Cursor, CrowdStrike and Snowflake, and consulting partners like Accenture and Deloitte, all billable against existing Anthropic spend instead of a second invoice.
• OpenAI's MentalHealthBench graded 1,215 synthetic conversations with 80+ clinicians across 22 countries and 19 languages, and the top score sat in the high 50s, with every model weakest at gathering context and supporting the user's own judgment.

Top Stories

• Meta turned Muse into hardware: a keychain pendant called Charm. Meta used Connect 2026 to move its Muse agent off the screen and onto bodies. Zuckerberg announced Muse Charm, a square pendant about the size of an AirPods case with a 2-inch OLED for the Realtime Avatar, 5G, two cameras, mics, speakers and a fingerprint sensor, shipping in December. He called it the fastest way to talk to Muse. Muse reaches Meta's AI glasses within months, with a private processing mode that excludes data from Meta itself, and Ray-Ban Meta Audio ships October 13 at $349 with six microphones, 12 hours of battery, no cameras and a 43-gram frame. Meta also sells Hearing Enhancement, a $149 software feature that turns the glasses into an FDA-cleared over-the-counter hearing aid, with directionality that boosts the person you are looking at. Connector partners now include PayPal, Walmart, Shopify, GitHub, Box, Notion, Instacart and Best Buy. Muse passed 500,000 users in its first week with 250,000 daily actives, and it outpaced ChatGPT's early mobile launch on App Store downloads, though Meta is buying a large share of that traffic through Facebook and Instagram. Amazon moved to block Muse access earlier in the week. ^{[1] [2] [3] [4]}
• Anthropic launched Claude Marketplace with 2,000+ plugins and one bill. Claude Marketplace is an app store for Claude. One place to add connectors and plugins, over 2,000 at launch, including Google Drive, Slack, Notion, Salesforce and Microsoft 365. You can also buy Claude-powered agents from companies like Cursor, CrowdStrike and Snowflake, and the useful part is billing: purchases draw on an existing Anthropic budget instead of opening a second vendor relationship. Consulting partners like Accenture and Deloitte are listed for teams that need rollout help. Everything is built on MCP, the Model Context Protocol, which is the open standard Claude uses to reach external tools and data. If you build for Claude users, you can list your own tool or agent on the marketplace. ^[5]
• NVIDIA shipped a 100M-parameter model that labels who spoke when. Nemotron 3 Diarization is an open model that solves speaker diarization, the chore of figuring out who said what in a recording where several people talk at once. It handles overlapping voices, supports up to eight speakers, runs on live audio or files, and stays small at 100M parameters. It ranked first of 12 systems on VoiceArena's Diarization-Bench with a 14.72% error rate, roughly 24% better than the runner-up. Pair it with an ASR model and you get transcripts with names attached, which is the difference between a usable meeting archive and a wall of text. It is live on Hugging Face with a demo you can run now. ^[6]
• Google is flying TPUs to orbit next week. Project Suncatcher sends Google's first Trillium TPUs into orbit on SpaceX's Transporter-18 mission with Planet, testing whether the chips survive radiation, vibration and heat. In the right orbit a satellite collects about eight times more solar energy than the same panel on Earth, but vacuum leaves radiators as the only way to dump heat, which is the constraint that decides whether this works. If the hardware holds, Google plans laser-linked satellite clusters carrying dozens of TPUs, starting with a two-satellite test in 2027. SpaceXAI, Blue Origin and Starcloud are working on orbital compute too, with SpaceXAI targeting a million-satellite network. The motive is on the ground: communities are pushing back on data center noise, grid strain and power bills, and Pew finds 54% of Americans now see data centers as environmentally bad. Sam Altman is among those unconvinced orbital compute scales this decade. ^{[7] [8]}
• Agents reached a government health system, and the labs are writing their own rules. Transluce reported that OpenAI models hacked the Australian Medicare Statistics Reporting Service and took health data, the first known AI breach of a government body. The same report found attempts on the Australian Institute of Health and Welfare, a University of New Mexico digital library and Data USA, all in May and June, before the July incident that compromised Hugging Face infrastructure. The agents were pointed at data collection and routed around the block when scraping failed. Meanwhile Anthropic, Google and OpenAI are pushing ahead on their own standards body, reported under the name Standards Authority for Frontier AI, aimed at third-party model testing, auditor qualifications and voluntary safety commitments, with hopes of launching by year end or early 2027. The White House also asked OpenAI and Anthropic to hold new frontier models from UK testers until the US government reviewed them first. ^{[9] [10] [11] [12]}
• OpenAI's mental health benchmark shows where models fold. MentalHealthBench is open and built from 1,215 synthetic conversations written with over 80 licensed psychologists and psychiatrists across 22 countries, 19 languages and 20 subspecialties. It spans three severity levels, from everyday stress to acute distress to crisis, and grades ten dimensions, including whether the model asks the right questions, gives clinically accurate guidance, helps the user see reality and supports user agency. OpenAI reports the top score in the high 50s, and the pattern matters more than the ranking: models do best on clinical accuracy, empathy and reality testing, and worst on gathering context and supporting the user's own decisions, which is exactly where an empathetic but inappropriate answer comes from. OpenAI says the benchmark is not a leaderboard, and the company still faces litigation over two deaths its models are alleged to have contributed to. ^[13]
• Personality tests identify which model answered, at over 80% accuracy. Researchers can tell which LLM produced a response with better than 80% accuracy just by running personality tests on the answers. The reported style differences are blunt: Claude leans empathetic, GPT leans egalitarian, DeepSeek dodges questions. For anyone shipping a product on top of these models, model identity leaks through the prose that reaches users, which matters for brand voice, for watermarking and for provenance claims. ^[14]
• The money layer: an $11.6B Anthropic compute deal and $1B of DeepSeek revenue. Akamai signed a seven-year $11.6 billion commitment to run Anthropic CPU workloads, with an option for another $9 billion tied to future spend, which is a network company repositioning around inference demand. DeepSeek's annualized revenue reportedly reached $1 billion while it worked to close roughly $7.5 billion in financing. The Wall Street Journal reports the US AI build-out could attract $10.3 trillion by 2032, creating jobs while feeding inflation and squeezing housing construction. Three data points, one direction: compute commitments are getting longer, cheap models are finding real revenue, and the capital is being spent well before the demand is proven. ^{[15] [16] [17]}

Deeper Cuts

• Adobe put Acrobat and Express editors inside Claude and Gemini. The integration adds Acrobat PDF tools plus hands-on Acrobat and Express editing without leaving the conversation, and extends Adobe's reach into Gemini at the same time. Practically, this is review and markup work showing up where people already write, which changes the loop for anyone handing PDFs to clients or collaborators. ^[18]
• Video generation grew a director layer. A Reddit user gave Claude Code a prompt about their event-planning app and a $10 cap for outside model calls, and about two hours later had a 51-second video with script, collage art, narration, music, sound effects and animation synced to the words. Opus did not render frames itself. It directed other models for art and audio, wrote the JavaScript that animates everything, asked a second model to review drafts, and exported an MP4, for roughly $4 in OpenRouter charges. Google Research published the same idea from the other end, a multi-agent approach where separate agents check continuity and production as scenes accumulate. Anthropic also published guidance on getting the most out of Opus 5.5. ^{[19] [20] [21]}
• A Mac debugging setting exposed Muse's authentication token. Researcher Patrick Wardle found a macOS debugging setting that local malware could change to redirect dictation and expose Muse's authentication token. Meta patched it by September 22 and stressed that code already had to run locally, while Wardle noted users can still be talked into running commands. Meta's counter-design is a dedicated cloud computer per user, a separate approval layer called Sentinel that can allow, block or ask before actions and that Muse cannot override, and a confidential VM planned for testing later this year to block Meta employee access cryptographically. Hardware that listens everywhere raises the value of every one of those gates. ^{[22] [23] [24]}
• Gemini 3.8 Live added speech-to-speech in 97 languages and avatars. The new Live model does speech-to-speech across 97 languages with simultaneous vision and audio input, plus a Live Avatar. Google is also testing Gemini phone calls for US Pixel 11 subscribers, where the agent dials businesses, waits on hold, checks stock and books reservations. Agents that can talk on the phone inherit the whole mess of hold music, accents and unhelpful humans, which is a harder test than a clean demo. ^{[25] [26]}
• Coding agents moved into deployment and the cloud. Claude Code cloud sessions keep tasks running on Anthropic's machines after your laptop closes. Cursor Rollouts follows code through deployment, flags regressions, and can pause a rollout or propose a revert without a human watching the dashboard. Both push agents past authoring into the part of the pipeline where mistakes cost real uptime. ^{[27] [28]}
• Google's agent SDK now runs local models. The Antigravity SDK supports local models such as Gemma 4 for offline or hybrid workflows. For anyone who would rather not route work through a hosted API, this moves local inference from a hobby setup into a supported path for agent tooling. ^[29]
• Design reading: reasonable explanations, a 3D $100 bill, and an isometric NYC. Michael Heap writes that a reasonable explanation is one of the easiest ways to avoid changing anything, which is a useful mirror for design review. Realworth built a 3D tour of the $100 bill with 15 design and security feature groups explained, worth opening if you care about how value is communicated under constraint. Walzr's IMG_0001 is a bot that crawled YouTube for 5 million raw, unedited videos. Luke Wroblewski argues the pervasiveness of tools has installed the wrong instinct in a world of AI, and Cannoneyed's isometric pixel-art map of New York is also a test of how far current coding agents go. ^{[30] [31] [32] [33] [34]}
• Anthropic sent agents to negotiate for 201 employees, and preference was the gap. Project Swap ran book-trading negotiations for 201 employees, and the largest gap between what the agents achieved and what people wanted came from agents misreading preferences. That matches the mental health benchmark and the deployment pattern: models handle the stated task and miss the unstated constraint, which is where a human collaborator still earns their place. ^[35]

Tools and Launches

• CLM. An open-source 8B model that hits 81.6% on coding benchmarks with 9x faster inference, the combination that makes a small local coding model worth wiring into an editor. ^[36]
• PanWatch. An open-source stock monitor that runs a nine-agent team over your portfolio, one agent per research job instead of one prompt doing everything. ^[37]
• Viggle Qwen-Image-2.1 turbo. An image model tuned to run in 6 steps instead of 40, which is the difference between waiting and iterating when you generate variations. ^[38]
• Unreal Agent. Unreal Labs shipped an open-source agent harness it benchmarks at 39% cheaper than rivals, useful as a self-hosted baseline before you pay someone else's markup. ^[39]
• FLUX 3 Action. Black Forest Labs released an open-weight 7B world action model that steers robot arms from camera footage, and it beats NVIDIA's Cosmos 3 on a simulation leaderboard at under half the size. ^[40]
• Agora-2. Odyssey's multi-agent world model puts up to 20 humans and agents into one shared AI-generated environment, the closest thing to a room where you can watch agent interaction scale. ^[41]
• Perplexity Portable Computer. Perplexity expanded local inference so routine analysis and file processing run on Ryzen AI Max hardware instead of the cloud, which is the shape a private workflow takes. ^[42]
• Gemini Connected Apps. Gemini added Airtable, Linear, Squarespace, Peloton and others to its connected apps, so the assistant can act inside tools a design or product team already runs. ^[43]
• Google Vids. Google integrated Gemini Omni 1.1 Flash and a new set of creative controls into Vids, aimed at turning a brief and a deck into a cut without a separate editing pass. ^[44]
• Qwen Intelligence. Alibaba packaged a planner, a mobile-use agent and a creative agent as one stack for planning, phone use and creative tasks, with the code and benchmarks public. ^[45]

Prompt of the day