Meta's next frontier model, still in training and internally codenamed Watermelon, has caught up with OpenAI's flagship GPT-5.5 on a set of closely watched AI benchmarks, chief AI officer Alexandr Wang told employees during an internal town hall on July 2, according to Business Insider and follow-up reporting from Benzinga and Bloomberg.
Wang's benchmark claim, with one big caveat
Wang, who runs Meta Superintelligence Labs after Meta acquired Scale AI last year, told staff that Watermelon has closed the gap to GPT-5.5 on the internal evaluations Meta uses to track frontier progress. He did not name the benchmarks. Meta has not published external evaluations and OpenAI has not commented, so the claim rests on one executive, one internal room and one round of unnamed tests.
An order of magnitude more compute than Muse Spark
Wang said Watermelon is being trained with an "order of magnitude" more compute than Muse Spark (previously codenamed Avocado), the first model in Meta's Muse series that launched in April 2026. He also flagged an upcoming Muse Spark update with stronger coding and agentic capabilities coming "pretty soon."
Why the timing matters
Meta has told investors it now expects to spend $125 billion to $145 billion in 2026 on chips, data centers and other infrastructure, up from a prior $115B-$135B guide, and has signed multi-gigawatt supply agreements with AMD and Nvidia. Wang's message to staff is essentially that those investments are starting to close the gap on OpenAI and Anthropic at the frontier - a claim shareholders will be watching once Watermelon actually ships.
Context: OpenAI and Anthropic keep moving
The claim lands the same week OpenAI is reportedly floating a 5% equity stake for the US government, and just after Anthropic shipped Claude Sonnet 5 and saw Claude Fable 5 return under lifted export controls.
Reporting based on coverage from Benzinga, Business Insider and Bloomberg.