Meta's most senior AI executive told staff last week that the company's next flagship model has closed the gap with OpenAI. The claim is striking. It is also, so far, impossible to check.
At an internal town hall in early July, Alexandr Wang, who leads Meta's superintelligence effort, said a model known inside the company by the codename Watermelon has "caught up" with OpenAI's GPT-5.5 on the benchmarks the field watches most closely. Wang did not say which benchmarks he meant, and neither Meta nor OpenAI has confirmed the comparison.
Bought with compute
What is clearer is how Meta got there. Watermelon is the successor to Avocado, the internal name for Muse Spark, which the company released in April. Training Watermelon has reportedly consumed an order of magnitude more compute than that earlier run. This is the brute-force route to progress: not a cleverer design but far more silicon and electricity pointed at the same problem.
The approach fits Meta's position. The company has spent heavily to assemble both talent and data centres, and Wang's own arrival was part of that push. A model that matches the field on raw benchmarks would be a return on the spending, at least on paper.
Why the claim needs a pinch of salt
Two things temper it. First, a number quoted at a staff meeting is not a published, reproducible evaluation, and the history of these announcements is littered with figures that softened once outsiders could test them. Second, the target may already have moved. OpenAI shipped GPT-5.5 in April and has since released GPT-5.6, which we covered when a watchdog flagged the newer model for benchmark gaming. Catching last spring's model is a milestone, but it is not the frontier.
None of this means Wang is wrong. It means the real test comes later, when Watermelon ships and independent researchers can run their own numbers. Until then the honest summary is that Meta believes it has caught up, and belief is not yet evidence.
Commentarii · 0