Word has emerged that the next AI model Meta is building has reached performance on par with OpenAI's GPT-5.5. Alexandr Wang, who leads the company's superintelligence group, is said to have made the remark at an internal all-hands meeting, in a report by Business Insider. The catch is that the benchmarks behind the claim have not been disclosed, and the model itself is still in training. Here is what we actually know about this bold performance claim.

What the "matches GPT-5.5" claim actually says

According to the reporting, Meta's next model, codenamed "Watermelon," has closed the gap with GPT-5.5 on widely followed benchmarks. Wang reportedly told employees as much at an internal town hall, and two people familiar with the matter shared the account with the press.

The point to watch is that it is not public which benchmarks he was referring to. There is no model card (the technical document summarizing performance and training conditions) and no independent third-party measurement. At this stage there is nothing outsiders can verify; it remains a self-assessment made inside the company. Given that the AI industry has repeatedly seen gaps between the numbers a developer reports and what outside labs measure, this warrants waiting for a public release and independent testing. Meta declined to comment on the report.

The difference from "Avocado" and an order of magnitude more compute

Watermelon is still in training, and no release date has been given. One characteristic reported about it is that it uses an order of magnitude more compute than its predecessor, codenamed "Avocado." Avocado is the model Meta released publicly this April as "Muse Spark," and Watermelon is the generation that follows it.

Increasing compute tenfold is not a minor bump. Training a large model requires stacking stages in order, starting with pre-training (the process of building the model's foundation on vast amounts of data), followed by fine-tuning and reinforcement learning; scaling that up sharply drives up both cost and the time training takes. If the claim is accurate, it can be read as a sign that the investment Meta has poured into chips and data centers is finally starting to show up as competitive results.

The backdrop: massive spending and a talent war

Meta expects to spend up to 145 billion USD (about 23.5 trillion yen) on infrastructure this year. Even so, the company has struggled to position its models at the leading edge of the industry. To change that, CEO Mark Zuckerberg brought in Wang last year and entrusted him with turning the effort around.

Meta has also thrown enormous sums at securing talent. According to reports, the company has offered top-tier researchers extraordinary packages worth hundreds of millions of dollars each (in the tens of billions of yen) to bring them on board. It is against this backdrop, where the wisdom of such spending is under scrutiny, that the internal "matched GPT-5.5" assessment has surfaced. Put another way, it is a remark made at a time of intense pressure to show results, and for that reason the claim deserves cautious reading.

*1 USD = 162 JPY (as of July 6, 2026)

The target itself keeps moving

Another thing not to overlook is that the point of comparison is not standing still. GPT-5.5 is the model OpenAI released this April, but the company rolled out a more powerful "GPT-5.6" late last month. That newer model, partly at the request of the US government, has not yet been made widely available.

In other words, the rival Meta says it has "caught up to" is already a generation old, and OpenAI's current frontier lies further ahead. Developing frontier models is a marathon in which the bar shifts the moment you catch up. Even if Watermelon genuinely has GPT-5.5-class ability, that does not mean it stands level with today's leading edge.

Summary

The claim that "Watermelon" matches GPT-5.5 points to the possibility that Meta's enormous investment is beginning to bear fruit, yet at this stage it does not rise above an internal self-report. The benchmarks are undisclosed, the model is still training, and there is no release date. GPT-5.5, the point of comparison, is already a generation behind. How Meta, which has assembled researchers with extraordinary pay and kept up spending in the trillions of yen a year, will demonstrate its results to the outside world remains to be seen. Until an actual model release and independent third-party evaluation are in hand, the truth of this claim cannot be judged.