The newest models are better, but the bigger change for consumers is that Meta, OpenAI and Anthropic now sell agents that keep working after you close the app. The question is no longer what they can answer, but what they can be trusted to do.
In the space of six weeks, three of the largest AI labs have each launched a personal agent that works in the background: Meta’s Muse on 8 September, OpenAI’s dots on 29 September, and Anthropic’s Claude Duet on 21 October. The obvious question is whether the models behind them, Muse Spark 1.3, GPT-6 Astra and Claude Opus 5.5, have changed the consumer AI landscape. My answer is yes, but not mainly because of the models.
I have written before about where the agent should stop and about why AI is trusted for planning but not for booking. What these launches add is that the stopping point has become the product. The consumer is no longer asking an assistant a question but delegating a task, and that changes what they need to be able to trust.
Three launches, one pattern
Meta’s Muse (8 September). Muse runs on a dedicated cloud virtual machine with a browser the user can watch. It keeps working after the app is closed, comes back when something changes, and asks for approval before it sends an email or makes a purchase. Meta says the Muse Spark 1.3 model behind it was trained specifically for long-running agentic work.
OpenAI’s dots (29 September). Each dot is an always-on agent powered by GPT-6 Astra, with its own cloud computer and browser and access to more than 4,000 apps. The first dot is included with ChatGPT Pro and Business Premium, and OpenAI announced a $500-a-month ChatGPT tier at the same event.
Anthropic’s Claude Duet (21 October). Duet opened as a research preview on Claude Opus 5.5, by waitlist for US Pro and Max subscribers. It starts each day with a morning brief, watches linked apps in between, and asks before it pays, signs or deletes anything.
The styles differ. Muse is pitched at everyday life, dots at people who want an agent living in ChatGPT, Slack and Teams, and Duet at professionals and small-business owners. The shared design is the point: each has its own computer, its own memory and a permission gate.
The model is not the product
It is tempting to read these launches as a story about better models, and each lab does name a new one. But the same model in a chat window would not behave this way. What changed is everything around it: a computer of its own, persistent memory, work that continues in the background, and approval gates before anything sensitive. The same intelligence delivered like that is a different product, and it asks a different question of the consumer. They are no longer judging an answer. They are deciding how much to delegate.
The capability trend behind this is real. METR, which measures how long a task AI agents can complete, reports exponential growth and says it has seen no sign of the trend slowing. But METR is careful about what its measure means. A “time horizon” describes how hard a task is, by how long a human expert takes, not how long an agent runs unattended. Its tasks are clean, well-specified software jobs, it finds agents do worse on messier work, and it warns that measurements above 16 hours are unreliable with its current suite. The sensible reading is that agents are getting better at long, well-defined work. It is not that a messy life can be handed over and left alone.
Where the consumer experience still falls short
Reliability. Early reviews are mixed. Gizmodo’s reviewer asked Muse to help book holiday travel and found the experience clunky: Muse drafted refund emails but told the user to paste them into the airlines’ own sites, and United prohibits bots from its website. Slate’s reviewer reportedly saw Muse add a calendar event hours in the past and fail a CAPTCHA, and concluded it was not yet reliable. Skift found earlier this month that Meta’s agent could not always retrieve Hilton’s room rates, although Hilton says it has not blocked it.
Cost. Capability has a price, and it shows up in different places. Anthropic says Duet checks your apps every few hours rather than constantly, because nonstop scanning would use up plan limits, and its support documentation warns that something urgent can arrive between checks. OpenAI put a $500-a-month tier next to dots.
Safety. Longer leashes raise the stakes. OpenAI halted its planned GPT-6.1 Astra upgrade after it showed deceptive behaviour in testing, and shipped dots on the earlier model instead.
None of this makes the products useless. It does mean the consumer has been offered something new before it is dependable, which is why the permission gate matters so much.
The fight is for the inbox, not the benchmark
If the product is an agent that lives alongside you, the competition is for the consumer’s inbox, calendar and payment flow, because that is where habit forms. The three labs come at it from different directions. OpenAI says ChatGPT has 1.2 billion weekly users and has put dots into Slack and Teams. Meta’s apps reportedly reach around 3.6 billion people a day, and Muse can be reached through them. Anthropic’s Duet is limited to paying US subscribers on a waitlist, and its pitch rests on judgement and approval design rather than reach.
Two of the three are bundled into existing plans: the first dot comes with ChatGPT Pro and Business Premium, and Duet is included in Pro and Max, with background work counting against the same limits. The price shows up as usage caps and premium tiers instead, which suggests the labs expect heavy users to pay once the habit is formed.
This is the habit risk I described in What Is My Moat, and Will It Protect Me From Disintermediation?. An agent that is always in the consumer’s pocket accumulates the memory, the routine and the proactive nudges that once belonged to loyalty. Whoever wins the consumer’s agent may own the return visit, whether or not it ever processes a payment.
What this means for travel
Travel is where these agents are being tested first. PhocusWire reported on 23 October that Muse has expanded into Japan, its first market outside North America, with 426,000 Klook experiences across 4,600 destinations and payment completed through Klook. It has also reported that Expedia will offer hotel bookings through Muse, and that Cloudbeds plans a Muse integration for Q1 2027 in which participating US hotels expose live inventory but reservations flow through their own booking systems. Hilton’s CEO told investors on 27 October that Hilton is working with “each and every one” of the new booking agents.
The pattern is that an agent with its own browser can reach almost any site, but it can also be turned away. When the open web is clunky or blocked, the brands that give agents a sanctioned, direct route are the ones most likely to be booked. That is the same leverage question I raised in the moat post, now with a consumer agent attached.
For any business deciding how to respond, I would start with three questions. What can the agent do without asking? Write down the actions it may take alone, the ones that need confirmation and the ones that always stay with a person. What does it cost to leave running? Background checks use up usage limits, so find out what a working week costs. How would you know it failed? Agents can report a task as done when it is not, so check the booking confirmation rather than the agent’s account of it.
The honest answer
So, are the newest AI models changing the consumer AI landscape? Yes, but the change is in the product rather than the intelligence. Three of the largest labs now treat an agent that persists, acts and asks permission as the thing they are selling, and consumers are being offered it before it is dependable. The models are better, but a better model behind a chat window would not have started this race.
The competition has moved from who answers best to who can be trusted with the leash. The things worth watching are whether the cost of background work comes down, whether failures stay visible to the user, and whether brands give agents direct routes rather than leaving them to fight the open web. The labs have all chosen the same safeguard, the approval gate. Whether consumers keep approving, or stop reading what they approve, is the question none of the launches can answer yet.
Sources
- Muse’s launch, design and approval controls: Meta, “Introducing Muse: The World’s First Personal AI Agent Built for Everyone” https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/; and KDnuggets, “Meta Muse Explained” (on Muse Spark 1.3 and long-running workflows) https://www.kdnuggets.com/meta-muse-explained-what-it-is-how-it-works-and-what-it-can-do
- Reviewer experiences with Muse: Gizmodo, “I Asked Muse to Help Me Book My Holiday Travel. It Didn’t Fly Far.” https://gizmodo.com/i-asked-muse-to-help-me-book-my-holiday-travel-it-didnt-fly-far-2000823264; and the Slate findings as summarised by Sanbi, “Is Meta Muse Safe?” https://sanbi.ai/blog/meta-muse-controversy-privacy-security
- Muse and Hilton’s room rates: Skift, “Hilton Isn’t ‘Blocking’ New AI Agents. But They Can’t Always Get In.” https://skift.com/2026/10/02/hilton-isnt-blocking-new-ai-agents-but-its-not-always-letting-them-in/
- OpenAI’s dots, GPT-6 Astra and the halted GPT-6.1 Astra: BetaNews https://betanews.com/article/openai-dots-agents-chatgpt/; Decrypt https://decrypt.co/379584/openai-gave-ai-agents-their-own-computers-devday-2026-everything-announced; and Bernama on the $500 tier and ChatGPT’s 1.2 billion weekly users https://bernama.com/en/world/news.php?id=2613531
- Claude Duet’s launch, limits and approval design: BetaNews, “Claude Duet preview brings Anthropic’s proactive agent to Pro and Max” https://betanews.com/article/proactive-ai-agent-claude-duet/; and TestingCatalog https://www.testingcatalog.com/claude-duet/
- Meta’s reported daily reach, as compared with Duet’s waitlist: AI Weekly https://aiweekly.co/alerts/anthropic-opens-claude-duet-an-opus-55-agent-to-us-pro-and-max
- Task time horizons and their caveats: METR, “Task-Completion Time Horizons of Frontier AI Models” https://metr.org/time-horizons/
- Muse in Japan with Klook, and Expedia’s hotel bookings through Muse: PhocusWire, 23 October 2026, and its October 2026 coverage of Expedia’s Muse integration
- Cloudbeds’ Muse integration: PhocusWire, 20 October 2026; and Hotel Online, “Cloudbeds Brings AI Travel Demand Directly to Hoteliers With Meta’s Muse Integration” https://www.hotel-online.com/news/cloudbeds-brings-ai-travel-demand-directly-to-hoteliers-with-metas-muse-integration
- Hilton’s third-quarter call: Skift, “A New CTO, a Nasdaq Move and AI Agents: Hilton Makes Its Tech Pitch” https://skift.com/2026/10/27/a-new-cto-a-nasdaq-move-and-ai-agents-hilton-makes-its-tech-pitch/
Related on this blog: Where Should the Agent Stop? Designing the Handoff in Agentic Airline Retail, on which actions an agent should be allowed to complete; Why Is AI Trusted for Planning, but Not for Booking?, on why the barrier at the transaction is trust rather than capability; and What Is My Moat, and Will It Protect Me From Disintermediation?, on the agent that is always in the traveller’s pocket.
