Meta's chief AI officer, Alexandr Wang, told employees this week that the company's next flagship model has drawn level with OpenAI's GPT-5.5, according to a July 2 Business Insider report sourced to attendees of an internal briefing. The model, codenamed "Watermelon" and still in training, is being built on what Wang described as an order of magnitude more compute than any Meta model before it. If accurate, it would be the strongest system Meta has produced, eclipsing both the Llama line and the Muse Spark family the company shipped in April.
It is also, for now, an unverified claim. Wang made the remarks in a non-public town hall. Meta has issued no official statement, disclosed no benchmark scores, and released no model. Reporting by Benzinga and American Bazaar, echoing the original account, noted that Wang did not specify which evaluations he was citing, and that neither Meta nor OpenAI has confirmed the comparison. Treat this as an early signal from inside a company with strong incentives to project momentum, not as a measured result.
What Wang Actually Said
According to the sourced accounts, Wang framed Watermelon as the successor to "Avocado" — the internal codename for Muse Spark, the model family Meta released publicly in April — and said it is currently in training. He told staff that Watermelon has "caught up" with GPT-5.5 on closely watched benchmarks and that it uses roughly ten times the compute of its predecessor. He also said a Muse Spark update with stronger coding and agentic capabilities would arrive "pretty soon."
The compute figure is where the story gets concrete. Meta's previous frontier run is widely estimated at around 100,000 H100-equivalents. An order-of-magnitude increase puts Watermelon in the neighborhood of one million GPU-equivalents — a scale that only recently moved from roadmap to reality. Meta is estimated to have roughly 600,000 H100s in production and announced plans earlier in 2026 for a one-million-GPU cluster. Analysts at SemiAnalysis have chronicled Meta's superclusters, Prometheus and Hyperion, as the physical substrate for exactly this kind of run. The claim is at least internally consistent with the infrastructure Meta has been pouring capital into.
The Strategy Behind the Fruit
Watermelon is the clearest test yet of Meta's expensive 2026 bet. Wang did not arrive at Meta by accident: he came in through Meta's roughly $14 billion investment in Scale AI, the data-labeling company he founded, a deal that also functioned as an acquihire installing him atop the newly formed Meta Superintelligence Labs. That reorganization was paired with a headline-grabbing talent raid on rival labs, with Mark Zuckerberg reportedly signing off on compensation packages large enough to reset the market for senior researchers.
The throughline is compute plus data plus people, applied at a scale few competitors can match. On data, Meta's advantage is structural: training material reportedly draws on proprietary corpora from Facebook, Instagram, WhatsApp, and Threads — a firehose of human interaction no lab can replicate. On compute, Meta is spending an estimated $125 billion to $145 billion on AI infrastructure this year. Watermelon is the output those inputs are supposed to justify.
There is a strategic tension underneath all of it. Meta built its AI reputation on open weights with Llama, positioning openness as both a moat and a differentiator against closed labs like OpenAI and Anthropic. It is not yet clear whether Watermelon will follow that path. A model trained on a million GPUs and proprietary social data is far more expensive to give away, and the superintelligence-lab framing leans toward frontier competition rather than ecosystem seeding. How Meta releases Watermelon — open, gated, or API-only — will say more about the company's real strategy than any benchmark claim.
What to Watch
The single most important development would be published, reproducible benchmarks, ideally third-party. Until then, "matches GPT-5.5" is a sentence from a closed room. Watch for a release timeline — none has been confirmed — and for whether Meta frames Watermelon as open-weight or closed. Watch, too, for the promised Muse Spark update, which lands sooner and will offer a nearer-term read on whether Meta's coding and agentic work is closing the gap that dogged its earlier releases.
Also worth tracking: how GPT-5.5 itself moves. Comparing an in-training model to a shipped competitor is a moving target, and by the time Watermelon is public the bar may have shifted again. Meta has assembled the compute, the data, and the talent. Whether that translates into a genuine frontier model — or another capable release that trails the leaders — remains, for now, an internal claim awaiting external proof.