Meta Platforms Inc. has accomplished a breakthrough in its bid to catch its artificial intelligence (AI) rivals: During a recent internal town hall meeting, Meta’s Superintelligence Chief, Alexandr Wang, informed employees that the company’s forthcoming AI model, codenamed Watermelon, has reached performance parity with OpenAI’s GPT-5.5 model based on closely followed internal benchmarks.
The disclosure, first reported by Business Insider, marks the clearest sign yet that Meta’s massive capital investments and aggressive talent acquisition strategies are starting to yield frontier-level results.
Watermelon is currently in its training phase and represents a massive scaling effort by Meta. Wang reportedly noted that the new model requires “an order of magnitude more compute” than its predecessor, Avocado, the internal codename for Muse Spark, which Meta released in April.
Muse Spark performed well on standard benchmarks, but Meta did not position it as a frontier-tier model capable of matching top-flight systems from OpenAI, Google, or Anthropic. Watermelon aims to change that narrative.
However, the specific benchmarks Wang cited to claim parity with GPT-5.5 have not yet been disclosed, leaving experts to view the milestone with cautious optimism until public data is available.
Despite the internal optimism, Meta leadership remains grounded regarding the steep hill left to climb. Meta CEO Mark Zuckerberg struck a cautious tone during the same town hall, admitting that Meta’s AI agent initiatives have not materialized as quickly as anticipated.
“The trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected,” Zuckerberg reportedly told staff, acknowledging that leadership had been overly optimistic about the development speed of tools like coding assistants.
What is more, should Watermelon successfully match GPT-5.5, Meta is still chasing a moving target. OpenAI debuted an even more powerful iteration, GPT-5.6, last month. While GPT-5.6 is currently restricted from public release at the request of the U.S. government, it signals that top-tier competitors remain a step ahead.
To close this gap, Zuckerberg has initiated a relentless spending campaign. Meta recently revised its infrastructure guidance, telling investors it expects to spend between $125 billion and $145 billion this year alone on chips, data centers, and specialized hardware — up from an earlier forecast of $115 billion to $135 billion.
Simultaneously, Meta has engaged in an aggressive talent blitz. Since hiring Wang away from Scale AI nine months ago to head the newly rebranded Meta Superintelligence Labs, the company has reportedly offered elite AI researchers compensation packages worth hundreds of millions of dollars to jump ship from rivals.
While Watermelon remains under wraps internally, Wang hinted at imminent, public-facing upgrades on social media.
In a recent post on X, Wang stated that an update to the current Muse Spark model is coming soon, promising major leaps in coding and autonomous agent capabilities. When questioned by a user on when Meta would field a coding model on par with Anthropic’s Claude Opus, Wang replied, “pretty soon,” adding that users would like what the company has “cooking.”
Meta declined to comment on the leaked town hall details, and OpenAI did not respond to requests for comment. Meta has not yet announced an official release date or open-source timeline for the Watermelon model.
