Happy Friday! The next place you meet an AI assistant might be your glasses. Meta is preparing to take Muse beyond the phone, Google is giving enterprise agents a face, and Anthropic is making its newest Claude cheaper to run. Here’s what deserves your attention—and a useful experiment to try this weekend.
🕶️ Meta wants your AI agent within speaking distance
Meta’s Connect announcements put its personal agent, Muse, at the center of a growing hardware lineup. The company says Muse will reach its AI glasses in the coming months, allowing the assistant to act on things you’re looking at, such as a product on a shelf or a flyer on a wall.
Meta also previewed Muse Charm, a pocket-sized device for talking to the agent, with more details promised later this year. New connections to retailers and work apps are part of the expansion.
Why it matters: Our take is that convenience could decide the next phase of consumer AI. An assistant you can reach while your hands are busy has a different job from one you open in a browser. The test will be how reliably it handles those everyday requests.
⚡ Claude Opus 5.5 makes a stronger case for affordable AI work
Anthropic released Claude Opus 5.5 on September 22, reporting improvements in coding and knowledge work alongside lower operating costs.
The headline number: 40% lower cost than Opus 5 on typical workloads in Anthropic’s tests. That combines lower token prices with greater efficiency. Standard input and output prices are $4 and $20 per million tokens, respectively—20% below Opus 5—while cached input is cheaper still.
Anthropic also reports faster output and stronger resistance to prompt injection, where malicious instructions try to redirect an assistant.
Why it matters: For teams running repeated research, coding, or document tasks, the useful measure is cost per satisfactory result. Our suggestion: compare models on the same real assignment, including the time you spend correcting the output.
🎭 Google gives Gemini enterprise agents an expressive face
Google introduced Gemini 3.8 Live with Live Avatar on September 24. Available through Gemini Enterprise, it pairs spoken conversation with a generated visual persona that responds with synchronized speech and expressions.
Google says the system can keep a conversation going while calling tools in the background, and supports synchronized speech and video across 97 languages. Custom avatars require enterprise allowlisting.
Why it matters: Our read is that customer service and guided product experiences are becoming a testing ground for more conversational interfaces. A convincing face may draw people in; solving their problem will determine whether they come back.
🧬 Research watch: Claude identifies an intriguing enzyme system
Anthropic’s new life sciences group reports that Claude agents identified previously uncharacterized features of an enzyme system while searching DNA sequence data. The campaign involved roughly 950 agents working for 21 hours, followed by analysis and experiments by human scientists.
The system, called array-associated reverse transcriptases, has repeat patterns reminiscent of CRISPR, the system behind widely used gene-editing tools.
The important limit: Its biological function remains unknown, and the findings are presented in a preprint. Our takeaway is the emerging research workflow: AI can help narrow a vast search space to candidates that scientists can investigate experimentally.
🗞️ Around the AI world
🎙️ ChatGPT brings connected apps into Voice. OpenAI’s September 23 update adds plugin support to Live on web and mobile. Voice in Work can also help create documents, presentations, and spreadsheets, subject to the account’s access and permissions.
🔗 Gemini adds another wave of connected apps. Google began rolling out integrations including Airtable, Linear, Adobe, Squarespace, and Webflow on September 23, expanding the work people can bring into a Gemini conversation.
🛠️ GPT-6 Sol and Luna arrive in Work and Codex. OpenAI introduced the models on September 22. Availability and reasoning options depend on the user’s plan and workspace settings.
🧪 Your weekend AI challenge
Pick one recurring research task and give an assistant with web access a clear, checkable deliverable. Try this:
Compare three tools for [my use case]. Use each company’s current official documentation and pricing pages. Create a table covering price, the features I need, the main limitations, and source links. Label anything you cannot verify. Recommend one option for [my budget and requirements], explain the trade-off, and finish with the three facts I should double-check before deciding.
Then inspect the links and record how much correction the result needs. That gives you a useful baseline for testing the next model release.
What would you actually delegate to an AI agent this week? Hit reply with one task you’d love to take off your plate.
Know someone who would enjoy this edition? Forward it and invite them to join AI Buzz!.