Meta just declared it’s no longer playing catch-up in the AI arms race. Alexandr Wang, the company’s chief of Superintelligence, told employees during an internal town hall on July 2 that the company’s latest model, codenamed Watermelon, has matched OpenAI’s GPT-5.5 on key benchmarks.
There’s a catch, though. The model is still in training, the benchmarks are internal and unnamed, and nobody outside Meta has verified any of it.
What we actually know about Watermelon
Watermelon reportedly uses approximately ten times more computational resources than its predecessor, the Avocado model. Avocado was part of Meta’s Muse Spark release back in April 2026, which dropped around the same time OpenAI launched GPT-5.5.
The specific benchmarks where Watermelon allegedly matches GPT-5.5 remain unnamed. That’s a meaningful omission. In AI, benchmark selection matters enormously. A model can look world-class on one set of evaluations while falling short on others. Without knowing which metrics Meta is referencing, the claim is essentially: “Trust us, we measured some things, and we did well.”
