Alexandr Wang, chief of Meta Superintelligence Labs, told an internal town hall that Meta's next frontier model — codenamed Watermelon — has caught up with OpenAI's GPT-5.5 on internal benchmarks, according to Business Insider's reporting from attendees this week. The model is still training. Wang didn't name the benchmarks. Meta declined to comment.
One detail in the pitch matters more than the parity claim: Watermelon is training on an order of magnitude more compute than Avocado, the internal codename for the model Meta shipped in April as Muse Spark. Ten times the compute, one generation apart. Wang also told staff a Muse Spark update with stronger coding and agent capabilities is coming "pretty soon."
The claim comes with an expiration date built in. OpenAI has already previewed GPT-5.6 — so by the time Watermelon ships, "caught up with GPT-5.5" may describe parity with last season's flagship. Catching a moving target by outspending it is a strategy; it's just not a cheap one.
Unverifiable parity is the new press release
Watermelon joins a pattern. The same week, xAI put Grok 4.5 into private beta at SpaceX and Tesla with Elon Musk claiming performance "close to, perhaps exceeding Opus" — also on internal evals, also with zero public benchmarks. Frontier labs have discovered that a confident town-hall claim generates the headline, and the system card can come later, or never. Until a third party runs the numbers, "matches GPT-5.5" is a recruiting and morale statement, not a measurement.
What is verifiable is the spend. A 10x compute jump per generation is exactly the trajectory that has Google's power bill up 37% in a year and Nvidia inventing financing structures to keep GPUs moving. It also explains why Meta is suddenly interested in renting out its machines: a fleet big enough to 10x your training runs is a fleet with idle capacity to sell, which is the whole logic of the "Meta Compute" cloud push reported this week. The training run and the cloud business are the same story — one enormous capex bill, two ways to justify it.
Our take: Wang's own framing is the tell — no clever architecture, just 10x the hammer. That's bullish for everyone selling hammers (chips, power, datacenter builds) and neutral-at-best for Meta until someone outside Menlo Park reproduces the benchmarks. Treat internal parity claims the way you'd treat a fund quoting its own unaudited returns. The number that can't be spun: compute per generation is now growing 10x, and that bill lands on capex guidance in three weeks.
What to watch
- Q2 earnings (late July): Meta's capex guidance is where a 10x training run stops being a codename and starts being a number.
- The Muse Spark update: Wang promised stronger coding and agent capabilities "pretty soon" — a shippable product claim, easy to check.
- A Watermelon system card: named benchmarks or third-party evals would separate measurement from morale. Silence past launch would say plenty.
- GPT-5.6's public release: if it lands before Watermelon does, the goalposts officially moved.