Meta Watermelon (the Unreleased Model): What We Actually Know & What We Don't

As of July 2026, Watermelon is not a model you can use. It is an internal codename for a large model that Meta Superintelligence Labs is still training, and the single performance signal in existence is one unverified thing an executive reportedly said at a town hall. This piece separates the verified facts from the hype, maps the confusing codename lineage, and gives you a reusable checklist for reading any vendor benchmark claim.

What we actually know

Start with the uncomfortable truth: almost nothing about Watermelon is measurable yet. It is an internal codename, not a product. Meta Superintelligence Labs (MSL) is reportedly still training it, and Meta has published no model card, no benchmark table, and no release date. Every performance sentence you read about it, including in this article, is downstream of a single scoop.

Here is the honest split between what is verified and what is not. Confidence levels are ours, based on how each claim is sourced.

The only hard, first-party numbers in this whole story are Meta's own capital figures (covered below) and the relative ten-times-compute framing. There are no published scores to reason about. If you see a headline treating Watermelon as a shipped GPT-5.5 competitor, it is compressing an unverified internal remark into a fact.

The GPT-5.5 claim, unpacked

The entire performance narrative comes from this: Meta AI chief Alexandr Wang (the former Scale AI CEO who joined Meta in 2025 to build MSL) reportedly told employees at an internal town hall around July 2 to 5, 2026, that Watermelon matches OpenAI's GPT-5.5 on some benchmarks. Business Insider originated that report; Benzinga, American Bazaar, Storyboard18, AOL and others re-reported it.

It is worth being precise about what kind of statement this is, because the dozens of republications can create a false sense of corroboration. They are not independent confirmations. They are all pointing at the same private meeting.

• The benchmarks were never named. No task, no dataset, no split. Parity on an unnamed benchmark is unfalsifiable.
• No configuration was disclosed. We do not know the checkpoint, decoding settings, or whether the numbers came from a favorable internal harness.
• There is no independent runner. Nobody outside Meta has evaluated the model, because the model is not out.
• The source is self-interested. The claim was made to employees during a year in which Meta raised capex and cut staff. That does not make it false; it makes it a claim that needs external evidence before it counts.
• The target already moved. OpenAI reportedly previewed GPT-5.6 shortly after. Parity with GPT-5.5 is measured against a version OpenAI has stepped past.

None of this means Watermelon is weak. It means the correct posture is agnosticism. For models you can actually run today, our best LLM of 2026 roundup and the GPT-5 vs Gemini 3 comparison are built on published, reproducible evals rather than town-hall quotes.

Codename lineage: Avocado to Watermelon

Coverage frequently muddles the fruit codenames, so here is the clean map. MSL uses produce names internally.

• Avocado is the internal codename for Muse Spark, which launched in April 2026 as Meta's first major model after Wang joined.
• Muse Spark 1.1 (agentic and coding focus, exposed via the new Meta Model API) shipped around July 9, 2026. This is the model you can actually use today.
• Watermelon is the codename for the next-generation successor, still in training, reportedly at roughly 10x the compute of Muse Spark.

Do not conflate Muse Spark 1.1 (shipped) with Watermelon (unshipped). They are different models at different stages.

How to read a vendor benchmark claim

This is the reusable part. Watermelon is just the latest case study in a recurring pattern: a lab makes a favorable comparison before anyone can check it. Run any such claim through this checklist. If it fails three or more, treat the number as marketing, not measurement.

The Watermelon claim fails five of six. That is the information gain here: not a guess at scores, but a decision rule. When a claim clears most of these, as published open-weight results usually do (see our best open-source LLM of 2026 piece), you can trust it. When it fails most of them, as this one does, you wait for the eval.

The money frame: capex and layoffs

Watermelon is as much a financial and morale narrative as a technical one, and the 2026 backdrop is hard to ignore. In the months leading up to the town-hall claim, earlier in 2026:

• Meta raised 2026 capex guidance to $125B to $145B (up from $115B to $135B) around its spring earnings, mostly for AI infrastructure, nearly double its 2025 capex of about $72.2B.
• Meta announced roughly 8,000 layoffs (about 10% of the workforce) in May 2026, attributed to funding that infrastructure.
• Zuckerberg reportedly acknowledged that Meta's AI progress had not accelerated the way the company expected over the prior few months.

Put those together and a plausible reading emerges: a bullish internal benchmark claim, made to employees during a year of cuts and a huge spend commitment, functions partly as a morale and investor story. We are not asserting that as motive, only noting that the incentive structure is real and should temper how much weight you put on an unverified parity statement.

What this means for builders

Practically, Watermelon changes nothing about your stack today, because you cannot call it. The right moves are unglamorous:

• Do not architect around it. There is no endpoint, no pricing, no latency profile, no release date. Building assumptions on an in-training model is building on rumor.
• Benchmark what ships. Muse Spark 1.1 is real and callable. Run it against your own task set next to GPT and Claude rather than trusting anyone's headline.
• Keep switching cheap. The whole reason vendor claims stress people out is lock-in. If your code speaks one OpenAI-compatible API, adopting Watermelon on day one (if it earns it) is a one-line model-string change.

That last point is why a gateway matters. For a broader model-by-model view of the current field, our Claude vs ChatGPT comparison covers the shipping frontier you can actually deploy right now.

FAQ