Meta Says Its Unreleased Watermelon Model Has Caught Up to GPT-5.5

Meta's head of AI, Alexandr Wang, told employees during an internal meeting that a model currently in training under the codename Watermelon has caught up to OpenAI's GPT-5.5 on select benchmarks. It's the first signal in months that Meta may be closing the gap with the leaders in the race for the most powerful language models, though the company itself has not confirmed any details.

The news was first reported by Business Insider, citing accounts from an internal meeting of Meta Superintelligence Labs, the unit created a year ago specifically to catch up with OpenAI, Google DeepMind, and Anthropic. Wang reportedly told the team that Watermelon, the successor to the Avocado model, is reaching a level close to GPT-5.5, OpenAI's most powerful model, which has been more widely available since April 2026.

What Wang Actually Said

According to accounts cited by U.S. media, Wang said Watermelon's results match GPT-5.5's level based on select benchmarks, but he gave neither their names nor complete figures. Meta declined to comment, and OpenAI did not respond to requests for comment. The absence of public, verifiable results makes the claim difficult to assess independently of the company's own assurances.

Wang also said an update to the Muse Spark model is coming, with major improvements to coding and agentic capabilities. Asked directly when Meta would release a coding model on par with Anthropic's Claude Opus, he only said it would happen soon and that users would appreciate what the company has in store.

Our next Muse Spark update is coming soon. Major improvements to coding and agentic capabilities. - Alexandr Wang, Chief AI Officer at Meta

Ten Times More Compute

A key element of the reports is the scale of resources needed to train Watermelon. The model reportedly uses roughly ten times more compute than Avocado, the internal name for the Muse Spark model released in April 2026. That's a massive jump, one that lines up with Meta raising its infrastructure budget for the year to $125-145 billion, compared with the previously planned $115-135 billion. The company is investing in chips, data centers, and supporting systems on a scale comparable to the industry's biggest players.

Such a gap between the resources spent and the claimed payoff raises questions about the efficiency of Meta's approach. If matching a model from several months ago requires ten times more compute, the progress curve relative to resources spent looks considerably less favorable than that of competitors achieving similar or better results with comparable or smaller investments.

Zuckerberg's Cautious Tone

Tellingly, at the same meeting Mark Zuckerberg struck a far more measured tone than Wang. Meta's CEO admitted that the company's AI development has not accelerated as previously expected. Those remarks come against the backdrop of earlier, high-profile layoffs in the AI teams, which some observers read as a sign of leadership's unease about the pace at which Meta is keeping up with the competition.

For more than a year, Meta has run a costly recruiting campaign, offering top AI researchers contracts worth hundreds of millions of dollars to build a team capable of catching up with OpenAI and Anthropic. The creation of Meta Superintelligence Labs under Wang, the former head of Scale AI, was meant to respond to earlier delays and underwhelming results from previous models in the Llama family.

A Moving Target and No Verification

Analysts note that catching up to GPT-5.5 may turn out to be an achievement that arrives months too late, since OpenAI has already rolled out the next generation, GPT-5.6, in the Sol, Terra, and Luna variants, though for now with limited access restricted to select partners at the request of the U.S. administration. In other words, even if Wang's claims hold up, Watermelon may catch up to a model that is no longer the market leader before it even reaches public release.

For Polish companies and developers using Meta's models in the cloud or locally, what matters most is whether and when Watermelon reaches public availability, and whether the promised Muse Spark update actually improves coding capabilities enough to compete with Claude Opus or GPT-5.5. So far no release dates have been given, and Watermelon remains in training.

The market reacted calmly to the reports, with Meta's stock posting only a modest gain, suggesting investors are treating internal, unconfirmed benchmark comparisons with caution. The AI industry has already seen cases where claims of catching up with the leaders turned out, in practice, to be premature or hard to verify once a model actually launched publicly.

Sources: Meta's Watermelon AI Claims GPT-5.5 Parity (techtimes.com), Meta's Upcoming Watermelon AI Model Matches OpenAI's GPT-5.5 on Key Benchmarks (benzinga.com), Meta's Watermelon AI Model Matches OpenAI's GPT-5.5 on Key Benchmarks (technobezz.com), Meta AI chief says Watermelon model has caught up to GPT-5.5 (americanbazaaronline.com)