Hatch: Meta Prepares to Launch Its A.I. Agent for the Masses as Early as September

Set Trending Topics as a preferred source on Google.

Meta is reportedly preparing the launch of an A.I. agent aimed at ordinary consumers. The project runs internally under the code name Hatch and is said to be scheduled for late August or early September. That is according to the U.S. trade publication The Information, which cites internal documents, and several outlets have since picked up the account. Meta itself has confirmed neither the name nor the features, the launch date or the pricing.

For Meta, this would be a clear step away from the chat assistant it offers today. Meta AI mostly answers within a conversation, while Hatch is meant to act on its own: take in a goal, choose the steps required, operate connected services and see multi-step tasks through to the end. Trending Topics reported in spring that Meta was working on an OpenClaw-style A.I. agent for the mass market.

What Hatch Is Supposed to Do

According to the reports, the agent was trained and tested in simulated versions of common web services, among them DoorDash, Etsy, Reddit, Yelp and Microsoft Outlook. Early prototypes are said to include a customizable dashboard on which the agent builds small tools of its own, such as a fitness tracker or a travel itinerary.

The capabilities under discussion include:

• researching products and services and comparing options
• drafting emails, forms, summaries and recommendations
• working through multi-step tasks across several connected apps
• retaining preferences and context beyond a single conversation
• tracking a goal over a longer period, prices or product changes for example
• asking for confirmation before purchases, bookings, messages or account changes

None of this is confirmed. It also remains open which interfaces Hatch would use to connect to third-party services and which partners will be on board at launch. A great deal hinges on that: an agent becomes useful once it can move between services, and that is precisely where it becomes harder to secure, because every service brings its own permissions and interfaces.

A second point tends to get blurred in the coverage: based on what is known so far, Hatch is a different thing from Meta’s agentic shopping tool for Instagram. One is designed as a broad consumer agent, the other as a commerce feature inside the app. Whether the two projects will converge later is unclear.

The Price: Up to $199.99 a Month

The reported pricing has drawn the most attention. According to the internal documents, Meta is weighing several tiers, with the most expensive one costing up to $199.99 per month and offering considerably higher usage limits in return. That would sit well above the Meta One subscriptions the company has been testing in markets such as Singapore and Guatemala, and it would look expensive next to the top tiers from OpenAI, Anthropic and Google as well.

The reason lies in the cost structure. Agents that navigate the web for minutes at a time consume many times the computing power of a single chat response. On the business side, Meta already charges by consumption: the WhatsApp Business Agent has cost $2 per million tokens since the start of the month, which works out to roughly 4 to 5 cents per typical exchange. As far as anyone knows, the consumer price list is still being worked out.

Which Meta Models Could Be Running Under Hatch

Meta has said nothing about which model powers the agent. Four candidates can be inferred from the reporting so far and from the portfolio of Meta Superintelligence Labs.

Muse Spark is the most obvious foundation for the launch. Early Financial Times reports on Meta’s personal assistant already described it as an application built on Muse Spark, the labs’ proprietary flagship. Meta recently opened the developer version of the model with a pointedly aggressive pricing strategy, which fits a product that has to handle a very large volume of requests.

Muse Code is the second candidate, though less as the face of the product than as the machinery behind it. With the coding agent, Meta is taking on Claude Code and OpenAI’s Codex; it reportedly builds on a newer Muse Spark generation and is the company’s most heavily tool-chain-trained system to date. That same ability, planning, calling tools, checking intermediate results and carrying on, is what a consumer agent needs as it works its way through order forms and inboxes.

Muse Glimmer would be the inexpensive part of the stack. The open 30-billion-parameter model runs locally as an always-on agent on Mac and PC and would be a fit for simple, frequently repeated steps, such as keeping an eye on a pricing page. An agent that pursues goals over days generates mostly routine load, and running that on the most expensive model would be hard to justify economically.

Watermelon, finally, is the model the reported premium tier most likely hangs on. It is not expected until October, which is after the planned Hatch launch. If both timelines hold, the agent would start out on the current generation and be upgraded later. Internally, Watermelon is said to have reached parity with GPT-5.5, achieved with roughly ten times the computing power of its predecessor. Those figures come from internal benchmarks and remain independently unverified.

What emerges in the end is likely to be a routing setup rather than a single model: simple steps on small, cheap models, hard ones on the frontier model. The tiered usage limits that reportedly separate the plans point in the same direction.

How This Fits Meta’s A.I. Strategy

Hatch is the most visible attempt yet to turn Meta’s enormous A.I. investments into revenue that comes from somewhere other than advertising. In the second quarter, more than 97 percent of revenue still came from the ads business, while capital spending ate up most of the free cash flow. Trending Topics analyzed that in detail recently.

Mark Zuckerberg laid out the logic himself on the most recent earnings call. He assigns the available computing power to three buckets: the existing advertising and recommendation systems, agents and APIs for businesses, and consumer agents for the roughly 3.5 billion people using Meta’s apps. He defended this year’s announced capital expenditures of $125 billion to $145 billion with the argument that it would be foolish to simply sell all of the compute and take the short-term profit.

The origin of the idea is striking. The model here is OpenClaw, the open-source agent built by Peter Steinberger that went viral last year. Meta barred internal use of OpenClaw earlier this year on security grounds, shortly before Steinberger left for OpenAI. The difference from the template is distribution: OpenClaw is available and offers more control, but it usually requires self-hosting or managed hosting. Hatch is meant to run where people already are, inside Instagram and WhatsApp.