Meta's upcoming artificial intelligence (AI) model, codenamed 'Watermelon', has reached performance comparable to OpenAI's GPT-5.5 on internal benchmarks, Meta's chief AI officer Alexandr Wang told employees during an internal town hall, Business Insider reported, citing sources.
According to the report, Watermelon is still in training and is expected to succeed Meta's previous frontier AI model, codenamed Avocado, which was released publicly earlier this year as Muse Spark, the company's flagship large language model.
Wang reportedly said the model caught up with GPT-5.5 on key benchmarks, but did not specify which benchmarks were used or disclose evaluation data.
The report added that Watermelon is being trained using around ten times more computing power than Avocado. Wang told employees that future iterations of the model are expected to deliver stronger coding capabilities and improved AI agent performance.
The report comes days after Meta chief executive Mark Zuckerberg told employees that the company's efforts to build AI agents were progressing more slowly than expected. Speaking during an internal town hall, reported by Reuters, Zuckerberg said that Meta underestimated the time required for its AI reorganisation and that executives had miscalculated how quickly the changes would translate into improvements in AI agent technology.
He added that the company expects to see stronger returns from its AI investments over the next three to six months as newer models are launched.
Meta is running Superintelligence Labs, the division established earlier this year to oversee the company's frontier AI work, to get a headstart in a space crowded by giants like OpenAI and Anthropic. Wang was appointed to lead the unit when the tech giant acquired his startup Scale AI.
