Meta Watermelon Model Nears October Deadline After Hatch Became Muse

From codename to product, minus one model

In late August, The Information reported from internal Meta documents that the company planned to ship two AI projects close together. The first was a consumer agent platform codenamed "Hatch." The second was a new flagship model codenamed "Watermelon," aimed at October.24 Seven weeks later, the first half of that plan has shipped, and the second is up against its deadline.

Hatch came out on September 8 as Muse, a standalone personal agent that runs on Meta's Muse Spark family of models.22 Watermelon has not been released. With less than three weeks left in October, whether Meta keeps that internal date says more about its standing among AI labs than anything else in its 2026 roadmap.

What the August leak got right, and what it didn't

Outlets that picked up The Information's scoop told largely the same story. Hatch was described as a consumer version of the OpenClaw agent concept, with a launch in late August or early September.4 It had been trained to work with services including DoorDash, Etsy, Reddit, Yelp and Outlook. Early prototypes showed a customizable dashboard of agent-built tools, such as a fitness tracker or a trip planner.6 The reports framed the project as part of Mark Zuckerberg's effort to earn back Meta's heavy AI spending and grow revenue beyond advertising.2

The timing turned out to be accurate. CNBC confirmed that the Muse app was the product built internally as Hatch.22 Reuters reported that it was modeled on OpenClaw and carried the Hatch codename inside the company. Meta's technical post mentions the codename only in a footnote.27

The headline price did not hold. The leak said Meta had weighed a premium tier of up to $199.99 a month.3 At launch, Muse came with a free tier and two paid plans, Power at $20 and Maximum at $100. Both buy more usage, not extra features.28 Meta's help center lists weekly limits of 500 million tokens on Power and 3 billion on Maximum. Zuckerberg has said the free tier covers up to 100 million tokens a week.21 Chief AI officer Alexandr Wang told Axios that most people should be able to stay on the free tier, and that the paid plans mainly exist to cover compute costs for heavy users.29

The clearest split in the coverage concerns which model would power Hatch. One business-focused write-up reported that the agent would launch on Anthropic's Claude Opus 4.6 and Sonnet 4.6 and move to Watermelon later.5 The shipped product points the other way. Muse runs on Muse Spark 1.3, the model Meta released six days before the agent.2827 The leaked documents themselves never said whether Watermelon would join the Muse line or ship as a separate system.24 On the evidence available, Meta chose to launch with its own model instead of waiting for a stronger one or relying on a rival's.

Why Watermelon carries the weight

Muse depends on a model family that Meta has been updating quickly but that is not yet in the top tier. Meta Superintelligence Labs released Muse Spark on April 8 as its first model. The release broke with the open-weights Llama approach: Spark is a proprietary, natively multimodal reasoning model.16 Updates followed through the summer. Muse Spark 1.3 arrived on September 2, with a context window above one million tokens. Meta said it needs roughly 20% fewer tool calls and 25% fewer tokens than version 1.2 for the same tasks.20

Those are efficiency improvements, not a jump in capability, and outside observers have noticed. One assessment of Meta's superintelligence push pointed out that no Muse model has an independent score on the ARC-AGI leaderboard.19 Even investors who defend the company tend to put it behind Anthropic and OpenAI on released models, and they say Watermelon will decide the question.35

The main specific claim about Watermelon is a strong one. Coverage citing The Information said internal benchmarks put the model level with GPT-5.5, reached with about ten times the compute of Muse Spark.1 Zuckerberg added detail in a long interview after the Muse launch. He described Muse Spark as built on a smaller pretraining run codenamed Avocado, with Watermelon as a larger run coming next. He said it would ship soon and conceded that Meta had misjudged LLM scaling the first time.36

That account matters for how to read the launch. Muse is Meta's agent design running on its smaller model, and Watermelon is the larger model meant to go underneath it. If the parity claim survives independent testing, Meta would have a large consumer agent and a frontier-level model at the same time. If it doesn't, Muse becomes a well-designed product running on second-tier reasoning.

The October clock

There are real doubts about the date. As of August, Watermelon was reportedly still in training with no external release date. A tenfold increase in compute usually means longer stabilization and safety testing.31 On a thinly traded Polymarket contract, the single most likely outcome in early October was no release by October 31. The total volume was less than $30,000, so the signal is weak.31 Other commentators expect a late-October release and treat it as the start of Meta's agent era, along with its Prometheus data center coming fully online.32

My view is that Meta has strong reasons to ship on time, and its record this year makes that plausible. The lab released Muse Image, several Spark updates, the Muse Code terminal agent, the open Muse Glimmer 30B and the Muse agent between July and September.15 Rivals have not slowed down either: OpenAI, Google, Anthropic and Meta all released new models in the first week of September.20 A flagship slipping into November would not be a crisis. But Zuckerberg has publicly said Watermelon is coming soon, and that raises the cost of a quiet delay.

One caveat runs through all of this. Every Watermelon detail so far, including the timeline, the compute ratio and the GPT-5.5 comparison, comes from internal documents or executive comments, not published evaluations. Labs often claim benchmark parity internally, and those claims often look different after release.33 Until Meta publishes technical details and outside testers run the model, its capabilities are unknown.10

The money behind the model

The pressure to ship a frontier model is ultimately financial. Meta has guided 2026 capital spending to $130 billion to $145 billion, up from $72.22 billion in 2025. In the second quarter, advertising brought in $59.36 billion of its $60.80 billion in revenue.15 Muse is the company's first serious attempt at earning directly from AI. It carries no ads at launch, and Wang has said Meta is looking at commerce as another source of revenue.29 Meta President Dina Powell McCormick has gone further, describing a possible fee on money the agent saves users, alongside subscriptions.18

The early adoption numbers support the bet, though only partly. Sensor Tower data cited by the Los Angeles Times put Muse above 2.5 million US downloads by September 23. Meta has not released its own figures.21 Meta then extended the agent to small businesses on September 29, with connections to Asana, Zoom, Intuit, Box, Canva and Slack, as well as to Meta ad accounts.37 Retailers have pushed back. Amazon cut Muse off from Amazon.com on September 20, while Shopify's Shop Pay was integrated a day later.19

Compute is what ties the model to the business. A model trained with about ten times Spark's compute will very likely cost more to run per task. Meta's usage-based pricing means its heaviest users will absorb more of that cost, but a free tier allowing 100 million tokens a week gets more expensive with each upgrade to the underlying model.21 Part of what Watermelon has to prove is whether Meta can offer frontier-level reasoning to a mass audience without wrecking the economics it just set up.

The reading

The August reports were correct about Meta's direction and timing, and wrong about the $199.99 price. They also left open the question that matters most: how good the model under the agent really is. Meta shipped the product first and is now under pressure to deliver the model. Muse showed that Meta can build and distribute an agent quickly. Watermelon will show whether it can compete at the frontier on its own models, and October 31 is the first test.