The premium tier is sold in units of capacity. Five to ten times the daily allowance of a free version, resetting each billing cycle, is what internal plans describe as Hatch Plus [9]. Nothing in that describes a better assistant. It describes a meter with a monthly cap, priced 25 times above the $7.99 chatbot tier Meta started selling three months ago [2].
The cost curve underneath is steep enough to explain the shape. Peter Steinberger, who wrote the open-source OpenClaw that Hatch answers, ran up $1.3m in tokens in a single month running agents at scale [7]. He was not ordering dinner. Still, at $199.99 a month, one bill that size equals the revenue from about 6,500 subscriptions [4], and a product like this does not fail on the average user. It fails on the tail.
Which is why the price may not be a price. Jeremiah Dillon, who does product marketing at Confluent and worked at Stripe and Google before that, wrote that he could believe $199.99 as Meta's average Opus API cost per user per month [11][12]. The build history gives that reading something to stand on: Hatch was developed on Claude Opus 4.6 and Claude Sonnet 4.6, with a plan to move it onto Meta's own Muse Spark models before launch [14]. If the swap lands, the margin is Meta's. If it slips, Meta is reselling somebody else's tokens at retail. Meta shipped Muse Spark 1.2 and the Muse Code terminal agent on 5 August and a 30-billion-parameter open-weight model five days later [16], so substitution is not fantasy, but Watermelon is only targeted for October and The Information could not establish whether it arrives inside the Muse family or under its own name [15].
What the subscriber buys is not capability OpenClaw lacks. OpenClaw already handles email, calendars and other services through messaging apps, on the user's own machine, for nothing, provided you will do the command-line setup and make your own decisions about models, credentials and permissions [6]. Meta is charging for the absence of that afternoon, plus the hosting underneath. Prototypes reach a bit further and show a dashboard of tools the agent has built for the user, a fitness tracker or a travel itinerary [5], which sustains a recurring bill better than errands do.
Then there is the second position. Meta is building a WhatsApp feature for connecting to agents that other companies build, with a limited rollout possibly starting this week [13]. Whatever Hatch settles on becomes the figure those agents are compared against, set by the owner of the channel they arrive through.
None of this moves the accounts. Meta's non-advertising revenue in the second quarter came to $1.44bn [5] against $59.36bn from ads [17]. Our reading in July was that Microsoft's AI spending turned into cloud revenue while Meta's turned into a cash-flow hole [19]. Hatch does not close that. It tests whether consumers will pay Meta directly for anything, which is a smaller question and the one that has to be answered first.
