Meta Watermelon Model and Hatch Agent Anchor Paid AI Push

What was reported

In late August, The Information reported on internal Meta documents describing two linked projects. The first is Hatch, a consumer AI agent platform the documents said would launch within weeks. The second is Watermelon, a new flagship model aimed at an October release36. The coverage also describes Hatch as the consumer version of OpenClaw, the AI agent project. It casts the plan as part of Mark Zuckerberg's effort to earn money from Meta's heavy AI spending and depend less on advertising35. Meta weighed a tiered subscription for Hatch, and its top tier could reportedly cost as much as $199.99 a month in exchange for higher usage limits24.

The reported product details matched across outlets. Early versions of Hatch were trained to work with DoorDash, Etsy, Reddit, Yelp and Outlook. They were built around a customizable dashboard of agent-made tools, such as a fitness tracker or a trip planner23. Meta reportedly trained the agent in a sandbox that simulates those services9. A Meta spokesperson declined to comment2. Several outlets also reported a separate WhatsApp feature that would let users talk to third-party AI agents, with a small test said to be starting almost immediately47.

It is now mid-October, the month Watermelon was meant to arrive. Release trackers updated in early October still list Muse Spark 1.3, released September 2, as Meta's newest model1527. The October deadline is therefore still open. Whether Meta meets it will show how much its internal roadmaps can be trusted.

Where the coverage diverges

The reports agree on the basic plan, but they split on what powers Hatch and how strong Watermelon is.

One account says Hatch launched on Anthropic's Claude Opus 4.6 and Claude Sonnet 4.6, with Meta planning to move it onto Watermelon over time9. If accurate, Meta's first paid consumer agent would have shipped on a competitor's model, which runs against the image of a self-sufficient AI lab. A later roundup cites Yahoo Finance coverage of internal claims that Watermelon matches GPT-5.5 on Meta's own benchmarks while using about ten times the compute of Muse Spark1. Neither claim has been confirmed by Meta. Other outlets point out that Watermelon has no published specifications, independent benchmarks or official announcement58.

The outlets also disagree on where Watermelon sits in Meta's lineup. Several said the documents did not show whether it would join the Muse family or stand alone237. One analysis argued that this choice matters. A standalone model would point to a new architecture, while a Muse variant would mean scaling up an existing line5. Some aggregator coverage got the basic facts wrong. One forum summary said Watermelon would ship alongside Hatch and that no pricing had been discussed10, which conflicts with the original reporting on both counts.

The Muse agent and the pricing gap

The biggest open question is whether Hatch has already launched under another name. Meta released a personal agent called Muse on September 8. It can shop, make bookings and fill in forms, and Amazon blocked it on September 2012. It runs on Muse Spark, is available only in the US, and comes in free, $20 and $100 monthly tiers19. That fits the timing and features reported for Hatch. No coverage directly confirms that Muse and Hatch are the same product, so the link is an inference.

If they are the same product, the pricing ended up well below the $199.99 ceiling. That suggests Meta chose to grow its user base before charging power-user prices. The distribution argument still applies. OpenClaw has to be installed on a personal computer, while Hatch was described as running inside Instagram, which has more than 2 billion daily users6. Amazon's block shows the main weakness of agents that work through other companies' sites: those companies can shut them out.

Meta's 2026 model roadmap

Watermelon caps a year in which Meta replaced its model strategy. Meta Superintelligence Labs released Muse Spark on April 8. It was a natively multimodal reasoning model, and it was closed-source, a break from Meta's tradition of open Llama releases13. Meta ended the open Llama line that month. The 2-trillion-parameter Llama 4 Behemoth teacher model was never released14.

After that, releases came quickly. Muse Spark 1.1 and a paid API arrived July 9. Meta said the model scored 88.1 on the MCP Atlas tool-use benchmark, ahead of Claude Opus 4.8 and GPT-5.5, though it trailed both on SWE-Bench Pro coding14. Muse Spark 1.2 and the Muse Code terminal agent followed on August 515. Muse Spark 1.3 came on September 2. It uses about 20% fewer tool calls and 25% fewer tokens than 1.2, at an unchanged price of $1.25 per million input tokens and $4.25 per million output tokens18. Its most notable changes are in behavior. It asks clarifying questions when a prompt is ambiguous and checks with the user before consequential actions. Meta held back a "max reasoning" mode for more safety testing20. All of these are useful traits in a model meant to power an agent that places orders and handles email.

On costs, the reported tenfold compute jump for Watermelon would add to an already tight GPU and memory market1. Meta still claims that Muse Spark matched Llama 4 Maverick with more than ten times less compute. No Muse model has an independent ARC-AGI score, so its standing against rivals is unverified12.

The open-weight track

The Stocktwits report on Hatch described a two-track strategy: closed commercial agents to take on the market leaders, and open-weight models to keep developers engaged4. The record supports that reading, up to a point. On August 10, Meta released Muse Glimmer, a roughly 30-billion-parameter model under the Apache 2.0 license11. It is a distillation of Muse Spark 1.2 and Meta's first fully open release since Llama ended14. The license drops the user-count limits and EU restrictions of the Llama license17.

Independent testing of Glimmer was less kind. Artificial Analysis gave it 35 on its Intelligence Index, versus 38 for Qwen3.6-27B, and measured an 82% hallucination rate14. Meta's roadmap promises an open-weights Muse Spark and larger Muse models around the fourth quarter20. Those weights would be the first open release from Meta's current flagship family. For now, the open releases are smaller models distilled from the closed flagship, which Meta keeps to itself.

The competitive field Watermelon enters

Watermelon is arriving during an unusually busy release season. OpenAI, Google, Anthropic and Meta each shipped a new model within a few days of one another in early September18. Anthropic then released Claude Opus 5.5 on September 22 at $4/$20 per million tokens and Claude Sonnet 5.5 on September 28 at $2/$10. It completed the 5.5 family with Claude Haiku 5.5 on October 7 at $0.10/$0.50 for prompts under 100,000 tokens24.

Google announced Gemini 4 Argon on September 30, but only vetted cyber defenders can use it so far, at an introductory $2/$10 against a standard $4/$2030. Google is also adding Claude models to the agent in its Gemini Business product, alongside Gemini 4 Argon and Gemini 3.8 Flash28. Open-weight competition is strong as well. Xiaomi's MIT-licensed MiMo-V2.6-Pro leads open models on the Artificial Analysis index30, and Mistral says Large 4 weights will ship by the end of October30.

Meta's advantage so far is price. Muse Spark 1.3 is the cheapest model in the leaderboard top five, at roughly $0.10 per million blended tokens20. A Watermelon at GPT-5.5 level would still trail the GPT-6 and Claude 5.5 generations, which have shipped since that comparison point. That is why the internal parity claim may have aged badly by the time the model ships.

The takeaway

The Hatch-and-Watermelon story is better read as a business plan than as a model launch. Meta has already built most of the pieces: a quickly updated closed flagship, a cheap API, a consumer agent with subscription tiers, and a limited open-weight line. Watermelon is meant to make that system competitive at the frontier. If it ships this month with independent benchmarks, Meta's reversal from Llama will look like a deliberate strategy. If it slips, or if reports that Hatch relies on Claude hold up, Meta will be selling agents built on a weaker model than its rivals have, and low prices alone are unlikely to make up the difference.