Meta Claims 'Watermelon' AI Matches OpenAI's Flagship GPT-5.5

The Shot Heard 'Round the Valley

The AI arms race just got interesting. In an internal town hall, Meta's chief of superintelligence, Alexandr Wang, dropped a bomb: the company's upcoming foundation model, codenamed 'Watermelon,' has achieved performance on par with OpenAI's formidable GPT-5.5. The claim, first reported by Business Insider, is based on internal benchmarks. It’s the clearest sign yet that Meta's colossal spending spree on compute and top-tier talent is starting to seriously reshape the competitive landscape.

For a company long seen as playing catch-up to OpenAI, Google, and Anthropic, this is Meta planting a flag. While its Muse Spark models, released in April, were respectable, they never quite broke into that top tier. Wang's announcement, though still unverified by the outside world, suggests the Meta Watermelon AI model is a different beast entirely.

It's a product of brute force. According to Wang, the new model uses an "order of magnitude more compute" than its predecessor, 'Avocado.' So this isn't some story about a clever new architecture. It's a story of what happens when you point one of the world's largest firehoses of capital and hardware at a single problem.

A Mountain of Chips and a Billion-Dollar Bet

You can't build a frontier model without a mountain of silicon. And Meta has been buying mountains. CEO Mark Zuckerberg has been blunt about his ambitions, stating earlier this year the company was on track to acquire 350,000 Nvidia H100 GPUs by the end of 2024, creating a total compute infrastructure equivalent to nearly 600,000 H100s. At nearly $30,000 a chip, that's a multi-billion dollar investment before you even build the data centers to house them. The company's projected infrastructure spending for 2026? A staggering $125 billion to $145 billion.

But hardware is just one piece of the puzzle. The other is talent. Meta's appointment of Alexandr Wang in June 2025 to lead its new Meta Superintelligence Labs was a power play. Wang isn't just another executive; he's the founder of Scale AI, the critical data-labeling firm that ballooned to a $29 billion valuation and made him the world's youngest self-made billionaire. The deal saw Meta take a 49% stake in Scale AI for $14.3 billion, effectively buying one of the sharpest minds in the business. Wang's presence lends serious credibility here. He knows what a state-of-the-art model looks like—his old company was instrumental in helping build them for everyone else.

The Open Question in a Closed Race

So, the big question: will Watermelon be open-sourced? Meta has always set itself apart from OpenAI with its Llama models, which, while not open-source in the strictest sense, were widely available. That strategy has been a cornerstone of the open vs. closed AI debate, building a massive ecosystem around Meta's tech. Releasing a model with GPT-5.5-level capabilities would be like throwing gasoline on the global AI fire.

Here's the catch. The immense cost of training Watermelon complicates that calculus. The sheer scale of the investment might just push Meta toward a proprietary, revenue-first model to get its money back. This is Meta's crossroads. An open Watermelon could cement its role as the champion of the developer community. A closed one signals a head-to-head commercial war with OpenAI and Google for enterprise cash.

Beyond Parity: The Push for Agency

Reaching parity on benchmarks is a milestone. It is not the finish line. The true test for these powerful new models is their ability to perform complex, multi-step tasks. In a recent post on X, Wang himself hinted that an update to Muse Spark is coming with "big improvements in coding and agentic capabilities." This means these models are designed to do more than just chat. They're becoming engines for autonomous systems, or AI agents that are a leap, not just a step, that can actually execute tasks for a user.

This pursuit of 'agentic' AI is the real prize, powering everything from a smarter assistant in Meta's Ray-Ban glasses to complex business automation. But as these models get more capable, they inevitably attract more scrutiny. The industry is already seeing how breakthroughs can run into a government buzzsaw, with regulators getting nervous about the risks. If Watermelon is as powerful as Meta claims, it's launching into a political minefield, not just a marketplace.

For now, this is just talk. An internal claim. The pressure is on Meta to release Watermelon and submit it to the unforgiving court of public opinion and third-party benchmarks. But the message from Menlo Park is clear: the race for AI supremacy is no longer a two-horse affair. A third heavyweight just stepped into the ring, and it's backed by a war chest of compute that is second to none.