The same week Mark Zuckerberg told Meta staff that the company’s AI bets “haven’t come to fruition yet,” his superintelligence chief Alexandr Wang gave them a very different message. Watermelon, the model Meta still has in training, has caught up with OpenAI’s GPT-5.5 on closely watched benchmarks. Business Insider reported Wang’s claim from two people in the room. Reuters heard a recording of Zuckerberg’s admission. Wang named a rival, a model, and a verdict, and skipped the one thing that would let anyone verify it, the benchmark itself.Most of the coverage ran with Wang’s line as proof that Meta had finally reached the frontier. Wang named a rival, a model, and a verdict, and skipped the one thing that would let anyone verify it, the benchmark itself.Wang claimed parity and named no benchmarkThe claim is single-sourced and internal. When reporters asked, Meta declined to comment, and OpenAI did not respond. The two people who described the town hall to Business Insider could not say which benchmarks Wang had in mind. All the reporting offers is a codename — “Watermelon” — the successor to Avocado, which is the internal name for the Muse Spark model Meta shipped in April. There is no model card, no evaluation harness, no ship date, and no table to back it up. Muse Spark scored well on standard tests and still trailed OpenAI and Anthropic overall, so “caught up on benchmarks” from the same team arrives with a track record attached.The missing table is the actual signalWhen a vendor publicly declares parity, the benchmark table typically accompanies the claim. OpenAI ran one with GPT-5.5, and Anthropic runs one with every Opus release. The table is what turns a claim into a result other people can validate and argue with. Wang skipped that step and made the claim to his own staff instead. An internal benchmark that nobody outside the room can replicate is just a morale score displayed as a number.An internal benchmark that nobody outside the room can replicate is just a morale score displayed as a number. This isn’t a critique of Wang’s honesty, but rather a comment on the artifact’s current state: a sentence rather than a score. It matters more because of who is making it and when. Meta reports earnings this quarter, and morale is raw after the roughly 8,000 layoffs and 7,000 reassignments Reuters tallied this year. Wang then followed the town hall by posting on X that a near-term Muse Spark update would bring big gains in coding and agentic work, with more promise than numbers to back it up.Ten times the compute to tie a model OpenAI already passedAccept the claim at face value, and it becomes harder to celebrate. Wang described Watermelon as requiring an order-of-magnitude more compute than Muse Spark to reach GPT-5.5. By the time Wang spoke, OpenAI had already moved to a limited preview of GPT-5.6, the successor to that model. Wang’s best case is Meta allocating that much compute to match a competitor’s system, which was surpassed weeks earlier…Wang’s best case is Meta allocating that much compute to match a competitor’s system, which was surpassed weeks earlier, all within a capital budget of $125 billion to $145 billion for this year. His narrative conveniently left that context out.The honest counter is that in-training models keep getting better, and a public Watermelon with a real model card and benchmark reports could still change the model-selection math, even for teams locked into OpenAI and Anthropic today. If that ships and the numbers hold, I am mistaken. Until it does, the only verified thing to come out of Meta’s town hall that week was Zuckerberg’s statement, not Wang’s, and it was the admission that the bets have not paid off yet.The post Meta says it caught OpenAI. One thing is missing. appeared first on The New Stack.