Key Takeaways
- Meta claims Watermelon matches GPT-5.5 on internal AI benchmarks.
- Watermelon reportedly uses 10 times more compute than Muse Spark.
- Independent testing is still needed to verify Meta’s benchmark claims.
Meta Watermelon AI model, codenamed Watermelon, has caught up with OpenAI’s GPT-5.5 on closely watched AI benchmarks, according to comments made by the company’s superintelligence chief, Alexandr Wang, during an employee town hall.
The claim, first reported Friday by Business Insider, has not been independently verified, and neither Meta nor OpenAI has confirmed the reported benchmark results.
Business Insider, citing two people familiar with the meeting, reported that Wang told employees Watermelon remains in training and uses “an order of magnitude more compute” than Muse Spark, Meta’s April 2026 model, internally known as Avocado. The report did not specify which benchmarks Wang referenced.
Meta declined to comment on the reported remarks, according to Business Insider. OpenAI did not respond to a request for comment.
Compute remains Meta’s focus
The reported comments suggest Meta continues to rely on increased computing power to improve the performance of its AI models. Muse Spark performed well on several benchmarks after its April release, but generally lagged behind leading competing models.
According to Business Insider, Wang said the Meta Watermelon AI model requires significantly more computing resources than its predecessor. The reported increase aligns with Meta’s broader investment in AI infrastructure, including billions of dollars spent on chips and data centers as the company expands its AI development efforts.
Investing.com, which republished the Business Insider report, noted it was not immediately clear which benchmarks Wang cited. Without published evaluation methods or datasets, the reported claim cannot be independently assessed.
Experts await public validation
The reported benchmark comparison remains an internal claim rather than a publicly documented evaluation. Meta has not released a model card, benchmark tables or technical documentation for the Meta Watermelon AI model, and the company has not announced when it will become publicly available.
Industry observers typically rely on published methodologies and independent testing to compare frontier AI systems. Until such evaluations become available, the reported benchmark results should be viewed as preliminary rather than conclusive.
Business Insider reported that Watermelon is the successor to Muse Spark and remains under active development. OpenAI released GPT-5.5 in April and introduced GPT-5.6 late last month, according to the report, raising the competitive stakes among leading AI developers.
The next major milestone will be whether Meta publishes benchmark data and whether independent researchers reach similar conclusions about Watermelon’s performance relative to GPT-5.5 and GPT-5.6.
