Meta says its next AI matches GPT-5.5 performance
What's the story
Meta's upcoming flagship artificial intelligence (AI) model, codenamed Watermelon, has achieved performance levels on par with OpenAI's GPT-5.5 in internal benchmark tests. The announcement was made by Alexandr Wang, the head of Meta Superintelligence Labs, during a company-wide town hall meeting. Though still in training, the Watermelon model has already matched several closely watched AI benchmarks set by GPT-5.5.
Project scale
Watermelon uses significantly more compute than Muse Spark
The development of Watermelon is a major step forward for Meta, which has invested billions of dollars over the past year into AI infrastructure, specialized chips, and top-tier talent. This investment is part of an effort to catch up with competitors like OpenAI, Google, and Anthropic. Wang revealed that "Watermelon uses an order of magnitude more compute than Avocado," the internal codename for Muse Spark, the first model in a new family of AI systems launched by Meta in April.
Model evolution
Muse Spark lagged behind OpenAI's models
While Muse Spark performed well on several benchmarks, it was generally considered to be lagging behind the most advanced models from OpenAI and Anthropic. Wang has also said that an updated version of Muse Spark would be launched soon with "major gains" in coding and agentic capabilities. This is seen as another step toward narrowing the gap with competing AI systems.
Hiring spree
Zuckerberg's cautious tone during the town hall
Meta CEO Mark Zuckerberg took a more cautious tone during the town hall. He admitted that AI agents had not "accelerated in the way" executives had hoped. The company has made AI its top strategic priority and is investing heavily in it. Reports suggest Meta could invest up to $145 billion in AI infrastructure this year, along with an aggressive hiring campaign for leading AI researchers with compensation packages reportedly worth hundreds of millions of dollars.