Key Takeaways:
Meta's next AI model has closed the performance gap with OpenAI's flagship GPT-5.5, marking a turning point in the company's years-long effort to compete at the industry's leading edge.
Meta Platforms Inc.'s upcoming Watermelon AI model has matched the performance of OpenAI's GPT-5.5 on key benchmarks, the company's superintelligence chief said Thursday, indicating that its $125 billion-plus infrastructure bet is yielding results.
"Watermelon, our next model after Avocado, is currently in training," Alexandr Wang, head of Meta Superintelligence Labs, said during an internal town hall, according to two people familiar with the matter. "Watermelon uses an order of magnitude more compute than Avocado."
Avocado was the internal codename for Muse Spark, the first model in Meta's new family released in April. While Muse Spark performed well on benchmarks, it did not match or exceed models from OpenAI or Anthropic's Claude. Wang cited closely followed AI model benchmarks to support his claim that Watermelon has caught GPT-5.5, though he did not specify which tests — such as MMLU, HumanEval or GPQA — were used. OpenAI released GPT-5.5 in April and debuted its more powerful GPT-5.6 in late June, though the US government has requested that OpenAI delay a general release over security concerns.
Meta told investors it expects to spend between $125 billion and $145 billion this year on chips, data centers and other infrastructure, up from an earlier forecast of $115 billion to $135 billion. If Wang's assessment holds, it would validate the strategy of Chief Executive Officer Mark Zuckerberg, who appointed Wang last year to lead Meta Superintelligence Labs and has offered top AI talent hundreds of millions of dollars each to join the effort.
The Talent and Compute Blitz
Zuckerberg has made closing the gap with OpenAI, Google's DeepMind and Anthropic a central priority. Wang oversees a team of elite researchers known as TBD, along with hardware initiatives that Meta has been quietly building. The company's compensation packages for top AI scientists have reached hundreds of millions of dollars, Business Insider previously reported.
The jump to Watermelon — requiring an order of magnitude more compute than Muse Spark — highlights that Meta is now deploying clusters rivaling the largest known training runs in the industry. Muse Spark itself required substantial computing resources across multiple data centers, and the escalation to Watermelon suggests Meta is now operating at a scale comparable to the largest clusters operated by Microsoft Corp. and Alphabet Inc.'s Google.
What It Means for the AI Race
If confirmed by independent benchmarks, Watermelon's performance would put Meta on par with OpenAI's most widely deployed model, though OpenAI has already moved ahead with GPT-5.6. The question for investors is whether Meta can sustain this pace: the company's infrastructure spending has more than doubled from two years ago, and the next generation of models will require even more compute.
For investors, the implications cut both ways. Meta shares could benefit from the perception that its massive capital allocation is paying off, narrowing the valuation gap with AI leaders. But the escalating cost of competition — Meta's $125 billion to $145 billion in 2026 infrastructure spending represents roughly 40 percent of its projected revenue — raises questions about when these investments will translate into earnings growth. OpenAI, meanwhile, retains a first-mover advantage with GPT-5.6, which has not yet been benchmarked against Watermelon.
Meta declined to comment on Wang's remarks. OpenAI did not respond to a request for comment.
This article is for informational purposes only and does not constitute investment advice.
