What Wang Said at the Internal Meeting
According to Business Insider, citing the content of an internal all-hands, Meta Superintelligence Labs head Alexandr Wang told employees that the next-generation flagship model, codenamed Watermelon, has already caught up to OpenAI's GPT-5.5 on the internal benchmarks the company cares about. The model is still in training, using "an order of magnitude" more compute than the previous generation Avocado—Avocado being the internal codename for the Muse Spark series Meta released this April. Wang also publicly teased on X that Muse Spark will soon get an update focused on coding and agent capabilities.
Several Question Marks on This Claim
To be clear: Wang didn't say which benchmarks; "caught up" comes from an internal rally at the vendor that most needs a win; and there's currently no third-party verification whatsoever. Since Llama 4, Meta's flagship model has been absent from the top tier for over a year, and the team recruited at great expense needs to prove itself both internally and externally at once. The claim from the internal meeting is a signal, but not yet a data point.
What's Worth Noting Is the Investment
More credible than "caught up" is the compute figure: an order of magnitude higher than Avocado means Meta has genuinely poured its 600,000-H100-class compute reserve into this round of training. Even if Watermelon doesn't reach the internally claimed level at launch, Meta's determination to get back to the frontier-model table is real. The release date hasn't been announced, and "caught up to GPT-5.5" should be treated as an unverified one-sided claim until there's third-party evaluation.