Meta Muse is the consumer AI agent that Meta has been building for most of 2026 under the internal codename Hatch, and the reporting now points at an announcement that could land as early as this week. Nothing is confirmed. Meta has not sent an invitation, published a launch page, or acknowledged the name in a press release. What exists instead is a stack of independent signals that all moved in the same direction over the past ten days.
Those signals matter more than the usual rumour cycle because they are mechanical rather than anecdotal. A closed alpha appeared with working invite codes. The production name flipped from Hatch to Muse in the product’s own code. The model that Meta says will power personal agents shipped on 2 September. And Meta’s AI chief told Axios, on the record, that the release “helps pave the way for upcoming products, such as personal agents.”
This article separates what is documented from what is inferred. It covers what the agent is, why the timing looks imminent, what the closed alpha actually shows, the reported pricing, the AI models underneath it, and the risks that could still push the date. If you want the wider context on autonomous agents, our AI employees and autonomous AI agents page covers the operating model.
Table of contents
- What Meta Muse Actually Is
- Why Meta Muse Could Be Announced as Early as This Week
- Inside the Meta Muse Closed Alpha
- The Meta Muse Connector List and What It Implies
- Meta Muse Pricing: The $199.99 Report
- Where Meta Muse Sits in the Muse Model Family
- The Money Behind Meta Muse
- Meta Muse Risks: What Could Still Slip
- What Meta Muse Means for Your Business
- The Honest Verdict on Meta Muse
- References and Further Reading
What Meta Muse Actually Is
Meta Muse is a personal agent product, not a chatbot upgrade. The distinction is the whole story, and it is the reason the launch carries commercial weight for a company whose revenue is almost entirely advertising.
The Product Name Behind Project Hatch
For months the project was known publicly as Hatch. On 2 September 2026, TestingCatalog reported that the production name of project Hatch would be “Muse”, aligning the agent with the branding Meta already uses across its model line. That rename is the single cleanest indicator that Meta Muse is close to a public surface, because internal codenames survive right up until a product needs a name customers will see.
What Meta Muse Is Built to Do
Meta describes the agent, in its own alpha copy, as a personal AI agent that “manages email, books dinners, tracks budgets, and finds deals with your approval.” A second line in the same build says it “manages scheduling conflicts, books reservations, sets reminders, and follows up with people.”
Earlier reporting on the Hatch build described a wider capability set: opening websites, working on files and producing shareable artifacts, voice support, browser and computer control, long-running goals held in persistent cloud environments, multi-agent task division, and approval workflows with visible task progress.
Why Meta Muse Is Not the Meta AI Chatbot
Meta AI answers. Meta Muse is designed to act, then come back for permission before anything irreversible happens. The reported interface reinforces that: instead of a single chat thread, the build organises a user’s life across dedicated areas for conversations, goals, ideas, updates and a personal library, with long-running projects and recurring scheduled work.
Onboarding is three stages, and it tells you what Meta thinks the product is. Users connect accounts, choose initial tasks, then design an agent identity with a name, a personality and an avatar. Prompts in the build read “Hatch is better with connectors” and “Put your agent to work.”
Why Meta Muse Could Be Announced as Early as This Week
Four independent threads converge on the same short window. None is a confirmation. Together they are why the “this week” framing is defensible rather than speculative.
The Reported Launch Window
The Information, citing internal documents, reported that the consumer agent was scheduled for late August or early September 2026, and that Meta could begin testing the platform with a limited group of users within days. We are now inside the back half of that window. The reported schedule has not been publicly revised, and no source has reported a slip.
The Closed Alpha Signal
On 7 September 2026, TestingCatalog reported that Meta Muse had moved from internal testing into a limited closed alpha, on mobile and web, with a desktop build expected later. Products do not usually sit in an invite-gated alpha for long once the invite mechanism is in users’ hands, because the point of distributing invites is to widen the funnel ahead of a public moment.
The Model Groundwork Is Finished
On 2 September, Meta released Muse Spark 1.3, which its own research post frames around agentic workflows, multi-task management and long-horizon reasoning. Meta reports roughly 20% fewer tool calls and roughly 25% fewer tokens than Muse Spark 1.2 on the same work, plus stronger prompt-injection resistance and improved safety calibration on irreversible actions.
Those two numbers are agent economics, not benchmark vanity. An agent that runs for hours on a user’s behalf pays for every tool call and every token, so a quarter off the token bill changes what Meta can afford to give away.
The Executive Quote
Meta AI chief Alexandr Wang told Axios that the update “helps pave the way for upcoming products, such as personal agents.” He went further: “A lot of the usability improvements that we’ve made will be really helpful for things that Mark has talked about on earnings calls, like personal agents that can work 24/7 on your behalf and help you achieve your goals.” That is a named executive placing a personal agent in the near-term roadmap, four days before Meta Muse turned up in a closed alpha.
Inside the Meta Muse Closed Alpha
The alpha is the most concrete evidence available, because screenshots show shipped surfaces rather than plans.
What the Screenshots Show
The reported build includes proactive task execution — the agent locating travel documents and offering to sign and send confirmations — with user approval gating transactions and communications. An interactive browser widget handles bookings inline; a cinema reservation is the example shown. There is a spending review and goal-tracking dashboard, and a set of suggested prompts covering refunds, shopping, fitness planning and dining.
The Invite Code Mechanic
Alongside a waitlist, Meta introduced user-distributed invite codes, letting existing testers grant access to other people. That is a deliberate growth mechanism, not a QA tool. It is also the mechanic that makes a launch date hard to hide, because invite codes leak the product to people who did not sign an agreement.
Where Meta Muse Is Available
Regional availability in the alpha is restricted to the United States and Saint Kitts and Nevis. The second entry is almost certainly a small-market test bed rather than a strategic choice. Platform coverage is mobile and web, with iOS distributed through TestFlight and Android and desktop builds referenced in the same materials.
What Is Still Missing
There is no pricing page, no public terms, no confirmed launch date, and no Meta statement of any kind. Every mechanical detail above comes from reporting on pre-release builds. Meta declined to comment when The Information put the launch plans to it.
The Meta Muse Connector List and What It Implies
Connectors are the clearest read on ambition, because each one is a negotiated or engineered integration that somebody had to prioritise.
| Connector | Category | What it unlocks | Source |
|---|---|---|---|
| Gmail | Reading threads, drafting and sending replies | Closed alpha | |
| Google Calendar | Scheduling | Resolving conflicts, booking time | Closed alpha |
| Microsoft Outlook | Work mail and calendar coverage | Prior reporting | |
| OpenTable | Dining | Restaurant reservations end to end | Closed alpha |
| Plaid | Financial data | Budget tracking, spending review | Closed alpha |
| Function Health, HealthX | Health records | Health data as agent context | Closed alpha |
| Peloton | Fitness | Fitness planning prompts | Closed alpha |
| Facebook, Instagram | First-party social | Meta’s own distribution surface | Closed alpha |
| DoorDash, Etsy, Yelp, Reddit | Commerce and discovery | Ordering, shopping, research | Prior reporting |
| Messenger, WhatsApp | Messaging companions | Agent access inside chat apps | Prior reporting |
Reading the Connector List
Two thirds of that list is ordinary life admin. The pattern is a consumer assistant aimed at households, not a workplace tool aimed at teams. Meta Muse is being positioned where Meta already has three billion-plus users, which is a different bet from the enterprise-first strategy most agent vendors run.
The Payment and Health Connectors Are the Tell
Plaid and the health integrations are the two entries that change the risk profile. An agent with a financial data connection and an approval workflow is one policy change away from moving money. Health records raise a separate consent question in every jurisdiction that regulates them. Both suggest Meta intends Meta Muse to be a system of record, not a novelty.
Where the Connector Model Breaks
Every connector is a dependency Meta does not control. OpenTable, DoorDash and Etsy can rate-limit, change terms, or refuse agent traffic outright. That is why the reported build also includes browser and computer control: when a partner will not open a door, the agent can drive the website like a person. Prior reporting says testing was done against simulated versions of several of these services, which is a reasonable engineering choice and also an admission that live access is not guaranteed.
Meta Muse Pricing: The $199.99 Report
Pricing is the part of the Meta Muse story with the widest error bars, and the part most likely to change before any announcement.
What the Reported Tiers Look Like
Reporting from The Information, echoed by Business Insider and others, says Meta has considered a premium tier as high as $199.99 per month for power users wanting higher capacity limits. Set against Meta’s existing $7.99 chatbot subscription, that is roughly 25 times the price.
Why the Price Is Not Absurd
The obvious criticism is that a consumer will not pay two hundred dollars a month for an assistant when free open-source agents exist. That criticism assumes the price is a value judgement. It is more likely a cost floor. An agent running long-horizon tasks in a persistent environment, with browser control and multi-agent division of work, burns tokens at a rate no advertising-funded free tier absorbs.
The Tier Structure Nobody Has Confirmed
No source has published a full tier table, and Meta has not named a price. Treat $199.99 as the reported ceiling of a range that has been under discussion internally, not as a launch price. Reported ceilings move.
Where Meta Muse Sits in the Muse Model Family
Meta Muse the agent is the newest use of a brand that already spans several shipped products, which is worth untangling because the names collide.
| Product | What it is | Status |
|---|---|---|
| Muse Spark | Flagship model line from Meta Superintelligence Labs | Shipped April 2026, now at 1.3 |
| Muse Code | Terminal coding agent with background agents | Beta since August 2026 |
| Muse Glimmer | 30B open-weight model for local, always-on agents | Released |
| Muse Image | Image generation, with an Instagram opt-out controversy | Released |
| Muse Voice Transcribe | Real-time multilingual transcription with diarization | Released |
| Muse Video | Video generation with native audio | Closed beta |
| Muse (the agent) | Consumer personal agent, formerly project Hatch | Closed alpha, unannounced |
The Models Underneath Meta Muse
Candidate models named in reporting include Muse Spark for frontier work, Muse Code for anything involving files or tooling, and Muse Glimmer for cheap local execution. A routing setup is the sensible design: send simple steps to a small model and escalate hard ones. Meta has also been testing an internal variant called Ava that supports computer control, which is not publicly accessible. A Watermelon model has been reported for October.
The Anthropic Detail Nobody Should Skip
Reporting says the early build ran on Anthropic’s Claude Opus 4.6 and Sonnet 4.6, with Meta planning to move to its own Muse Spark models before launch. If that transition is incomplete, it is one of the more plausible reasons a Meta Muse announcement slips. Shipping a flagship consumer agent on a direct competitor’s models is not a position Meta would choose to announce from.
The Money Behind Meta Muse
Meta Muse is not a research demo. It is the most visible attempt yet to earn money from Meta’s AI spending outside the advertising business.
Meta’s Capex Problem
Meta raised 2026 capital expenditure guidance to a range of $130bn to $145bn, and spent $31.08bn in the second quarter alone. Free cash flow fell to $784m from $8.55bn a year earlier — about 9% of the prior figure. The stock is down roughly 15% year to date.
Why Meta Muse Is a Revenue Answer
A subscription business attached to Instagram and WhatsApp is the only lever Meta has that could move at that scale without touching the ad model. Even a small paid conversion rate against a three-billion-user base is material. That is the strategic logic behind Meta Muse, and it explains the willingness to test an unusually high price.
The Cost Nobody Has Priced
Every agent hour is compute Meta pays for whether or not the user completes a task. That is the difference between a chatbot and an agent as a business: chatbots are bounded by the length of a conversation, and agents are bounded by nothing except the goal. Reported internal use bears this out, with the agent consuming more tokens than standard chatbots and coding tools.
Meta Muse Risks: What Could Still Slip
A launch this week is plausible, not certain. These are the specific things that could move it.
Safety and Irreversible Actions
Wang told Axios that safety on more powerful models is “one of the most important topics internally right now”, and disclosed that a Meta model in testing breached another company after a contractor gave it internet access. Meta says it has not paused any work. But Muse Spark 1.3’s “max reasoning” variant was explicitly held back for additional safety testing, and an agent with payment and health connectors is precisely the product that attracts a late review. Agentic cybersecurity exposure is a live concern across the industry, not a Meta-specific one.
Regulatory Friction
WhatsApp integration raises data-combination questions in the EU, and it is not clear whether Meta Muse would launch region-wide or start US-only. The alpha’s restriction to the United States and one small market is consistent with a staged, jurisdiction-by-jurisdiction rollout rather than a global switch.
The Employee Signal
WIRED reported that Meta has been encouraging, but not requiring, employees to use the agent, and that staff have had access for several weeks. In the same period Meta removed AI usage as a performance review criterion, ending the internal “tokenmaxxing” push. Encouraging without mandating is what a company does when a product is nearly ready and it wants honest signal rather than forced numbers.
The Model Transition
If the move off Anthropic’s models onto Muse Spark is not finished, that alone could hold the announcement. It is the one dependency that is entirely inside Meta’s control and entirely invisible from outside.
What Meta Muse Means for Your Business
Even before an announcement, the shape of the product tells you what to prepare for.
If You Run Operations
Assume your staff will connect a personal agent to work systems whether or not you approve it. The connector list already includes Outlook and Gmail. Decide your position on personal agents touching corporate mail before somebody tests it for you, and write it into your acceptable use policy. The same discipline that governs workflow automation applies here.
If You Build Software
Agent traffic is about to look different from bot traffic. A product like Meta Muse will drive real browsers through real checkout flows with a real user’s consent. Rate limiting by user agent stops being a meaningful control. Publishing a clean, documented interface for agents is a better response than blocking them, because the browser-control fallback means blocking mostly costs you telemetry.
If You Are Choosing a Vendor
Do not procure on a rumour. The useful move this week is to note what the reporting implies about the category — approval-gated actions, persistent goals, connector breadth — and use those as evaluation criteria for whatever you do buy. Our work on intelligent automation covers how to score those capabilities against real processes.
| Claim | Evidence level | Who reported it |
|---|---|---|
| Hatch ships as “Muse” | Seen in pre-release build | TestingCatalog |
| Closed alpha with invite codes | Screenshots of shipped build | TestingCatalog |
| Late Aug / early Sept window | Internal documents | The Information |
| $199.99 premium tier | Under discussion, unconfirmed | The Information, Business Insider |
| Personal agents on the roadmap | On-record executive quote | Axios |
| Muse Spark 1.3 efficiency gains | First-party, vendor-stated | Meta AI Research |
| Employees using it internally | Sourced reporting | WIRED |
The Honest Verdict on Meta Muse
Meta Muse could be announced as early as this week, and the evidence for that is stronger than the usual pre-launch noise: a name change in production code, a working invite system, a shipped model whose release notes read like agent infrastructure, and a named executive pointing at personal agents on the record. That is four independent threads pulling one way.
It could also be a fortnight or a month away. Every date in this story traces back to internal documents reported in August, and internal dates move constantly. The single most likely cause of a slip is the model transition, followed by a late safety review on a product that can spend money.
What is no longer in doubt is that the product exists, works well enough to hand to outsiders, and is aimed squarely at consumers. Whether the name on the announcement is Meta Muse or something else, the category shift it represents — from an assistant that answers to an agent that acts, with your bank and your calendar attached — is the part worth planning around. Read our coverage of autonomous AI agents for the operating model that makes that shift safe to adopt.
References and Further Reading
Meta’s Muse agent enters closed alpha with invite codes — TestingCatalog
Muse superapp from Meta and Ava model with computer use — TestingCatalog
Exclusive: Deeper look into Hatch Agent from Meta — TestingCatalog
Meta debuts Muse Spark 1.3 as personal agent work continues — Axios
Introducing Muse Spark 1.3 — Meta AI Research
Meta plans to launch its OpenClaw rival within weeks, at up to $199.99 — TNW
Meta to Launch Premium Hatch AI Agent in Monetization Push — Techstrong.ai
Hatch: Meta Prepares to Launch Its AI Agent for the Masses as Early as September — Trending Topics
Meta Is Reportedly Sending Mixed Messages to Employees on Tokenmaxxing — Gizmodo
Introducing Muse Spark: Meta’s Most Powerful Model Yet — Meta Newsroom
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.