Meta ships Muse: a consumer agent inside a sealed VM
A consumer agent launch from Meta, an image model from OpenAI that halves its own latency, and a $32 million bet that robots can be tested in software. Tuesday's late run.
Meta launched Muse, its first consumer AI agent, and the entire pitch rests on one architectural claim: it runs inside a dedicated secure virtual machine that holds both the agent and the user's data. The agent is US-only at launch and reachable either through a standalone Muse app or by messaging it in WhatsApp. It handles both one-off chores — sending an email, booking travel — and long-horizon goals like building a yearlong training plan or setting up a new business, and Meta says it will open a browser, fill in forms, and negotiate on your behalf once you've handed it a goal. The connective tissue is the integrations: email, calendars, payments, health and fitness apps, smart home, dining, shopping, music, and events, plus your Instagram and Facebook accounts including the business and ads tooling. On price, Zuckerberg is making it hard to compete — up to 100 million tokens a week free, with paid tiers at $20 and $100 a month.
The privacy story is the whole ballgame, and it's a deliberate one. Zuckerberg told Alex Heath he personally helped recruit Signal's founder to work on a confidential virtual machine designed so that even Meta cannot see what Muse is doing for you, and Meta says Muse has no visibility into passwords or payment methods and doesn't feed conversations into its ads systems. That is a strong promise from a company that settled with 29 states for $18 billion in August over youth harms to children and was ordered to pay $942 million in a New Mexico case this summer. The technical work looks serious; whether it matters depends on whether people believe it. We covered the earlier piece of this stack in Meta ships Muse Voice Transcribe, its first real-time audio model — Muse has been assembling the senses for months, and this is the product they were for.
OpenAI released ChatGPT Images 2.5, and the headline number is a latency cut of up to 50% against Images 2.0. The model is better at holding onto the subject of a reference photo across new settings and styles, and better at editing only the part you asked for — both of which matter far more in production than raw prettiness. In ChatGPT, three features ship alongside it: Sketch, which turns a rough drawing into a real image; templates for common formats like flyers and product shots; and the ability to drop comments directly onto an image to target an edit. OpenAI says people now generate more than 3 billion images a week across ChatGPT Images and the API.
For developers the release splits in two. GPT-Image-2.5 Flare is the default — OpenAI claims higher quality than GPT-Image-2 at half the latency — while GPT-Image-2.5 Sunburst trades longer generation times for tighter control on detailed creative work. Early customers are the interesting signal: Adobe says it is putting both into Firefly, and Manus reports Flare running two to four times faster than GPT-Image-2 with better transparent backgrounds. Image generation is quietly moving from demo to dependency, and "edits only what you asked for" is exactly the property a design pipeline needs before anyone will wire a model into it.
Antioch raised a $32 million Series A led by Greylock to move robot testing out of the physical world and into continuously calibrated simulation. The 16-month-old company, founded by a former Tesla Autopilot engineer alongside ex-DeepMind and Meta Reality Labs researchers, builds what it calls a verifier: a digital twin of a customer's actual hardware, sensors, and operating constraints that predicts whether a proposed change will help before anyone touches a machine. A team that could afford one physical experiment can now run thousands of parallel scenarios, which is the same leverage software gets for free. Nvidia is a partner — Antioch is building on Omniverse, Isaac Sim, and Isaac Lab — and Nebius supplies the cloud.
The claim worth testing comes from Amazon, where Jason Mitura, a vice president of software development and chief product officer at Ring, says Antioch's simulations "closely matched our physical test results, including in scenarios we deliberately held out of calibration." Held-out scenarios are the right benchmark, because a simulator tuned until it matches your test data proves nothing. The open risk is commercial as much as technical: heavy per-customer calibration is how simulation companies accidentally become consultancies. Greylock is betting that the evaluation layer between models and machines becomes a control point in its own right — the same thesis Applied Intuition has been executing on in vehicles for years.
What to watch: whether Amazon or another named customer publishes sim-to-real transfer numbers, and whether Meta's sealed-VM claim gets an independent security review before Muse leaves the US.
Would you hand an agent your email, calendar, and payments if the vendor swore it couldn't see any of it? Tell us in the comments.
