Meta's in-training Watermelon model reportedly matches OpenAI's GPT-5.5 on internal tests

Meta Superintelligence Labs chief Alexandr Wang reportedly told employees at an internal town hall on July 2 that Meta’s in-training model, codenamed Watermelon, has matched OpenAI’s GPT-5.5 on internal benchmarks. Neither company has confirmed the claim.

The reported parity, if it holds up outside Meta’s own testing, would put the company back in contention at the frontier after a bruising stretch. Wang reportedly said Watermelon uses an order of magnitude more compute than Avocado, the model Meta released internally in April, a jump that lines up with the company’s stated plan to spend up to $145 billion on AI infrastructure this year.

The benchmark figure rests on a leaked account of a closed-door meeting, not a published result. Neither Meta nor OpenAI has said which tests were run or released any numbers, so the comparison cannot be independently verified. Internal benchmarks are also chosen by the team running them, which limits what a single “matches GPT-5.5” line reveals about real-world performance.

The upbeat model claim landed alongside a blunter admission. At the same town hall, CEO Mark Zuckerberg reportedly said AI agent development “over at least the last four months hasn’t really accelerated” the way Meta expected. He also reportedly conceded that a May restructuring, which cut about 10 percent of Meta’s global workforce and moved roughly 7,000 employees into AI teams, had not gone as cleanly as planned.

Zuckerberg reportedly said he expects more tangible AI results within three to six months, a timeline that turns the coming quarter into a test of whether the compute buildout is translating into shipped products. For now, the story is a company telling its own staff it has caught up, on tests only it has seen.