The model has a fruit name and a budget to match. Meta Platforms’ next-generation large language model, code-named Watermelon, has matched OpenAI’s GPT-5.5 on key benchmark scores in internal testing, according to people familiar with the matter, after Meta devoted roughly 10 times the computing power of its previous generation to the training run.
Meta’s head of AI disclosed the progress in an internal briefing, and Business Insider first reported the details. People close to Mark Zuckerberg said the chief executive views Watermelon as the decisive campaign in Meta’s effort to catch OpenAI, which has set the pace of the frontier model market since the release of ChatGPT.
The scale of the training effort is the headline. Spending ten times the compute of an earlier generation is not an incremental step; it implies a data center build-out and an energy bill that would strain the budgets of most technology companies. Meta has been expanding its AI infrastructure for two years, adding clusters of Nvidia GPUs and its own custom silicon, and Watermelon appears to be the first model to consume that capacity at full scale.
The benchmark parity is significant but qualified, researchers who have seen the results said. Watermelon reportedly matches GPT-5.5 on the standard reasoning and language tests that the industry uses as a rough scorecard. Those tests measure capability in controlled conditions, not behavior in products, and several recent releases have shown that benchmark leadership does not automatically translate into user preference.
Meta is also updating Muse Spark, its agent-focused model line, with stronger autonomous coding and tool-use abilities. The company is running two tracks at once: a frontier model to compete at the top of the market and smaller models tuned for specific jobs, from code generation to customer service. The dual approach mirrors a broader industry split, as companies decide whether to pour capital into ever-larger models or into smaller ones that can run cheaply at scale.
The Watermelon results arrive at a sensitive moment for Meta’s strategy. The company has published its previous flagship models, Llama and its successors, with open weights, a bet that free distribution would build a broad ecosystem around Meta’s technology. Whether Watermelon follows that path is a decision that will shape the economics of the release. An open model matches OpenAI’s closed products on paper but gives Meta no direct revenue from the model itself; a closed model would mark a departure from the strategy that made Llama the standard for open-weight AI.
Analysts said the more important number may be cost. Ten times the compute of a prior generation suggests training expenses that run into the billions of dollars, a bill that raises the bar for every future release. “Meta is effectively committing to a cadence of frontier training runs at a scale only a handful of companies can afford,” said one analyst who tracks AI infrastructure spending.
The stakes extend beyond the model itself. Meta’s advertising business depends on recommendation systems that are increasingly powered by large models, and the company has said AI agents will become the next interface for commerce on its apps. A frontier model that matches the industry leader gives Meta cover on multiple fronts: it keeps OpenAI from defining the technology, it powers in-house products and it signals to investors that the company’s massive capital spending is producing results.
Mr. Zuckerberg’s framing, according to people familiar with his thinking, is competitive rather than academic. He has described the gap with OpenAI in terms of a race in which second place in benchmarks can still win in distribution. Meta reaches more than three billion people through its apps, and the company’s plan is to route AI features through those properties rather than through a standalone assistant.
The compute commitment also has a physical dimension. A training run on this scale requires power purchase agreements, new substations and data centers planned years in advance, and Meta has been signing up electricity and sites accordingly. Rivals reading the Watermelon numbers are reading those contracts as a statement of intent: Meta is not dabbling in frontier training; it is building the permanent capacity for it.
The timing of the disclosure matters as well. OpenAI has been pricing its flagship products at a premium, and a credible competitor with a similar benchmark profile gives corporate buyers a second option at a moment when AI budgets are under review. Enterprise customers, analysts noted, have been slow to commit to single-vendor AI stacks, and Watermelon’s reported results give procurement teams a reason to wait.
The practical test comes later this year, when Watermelon is expected to reach products. Internal benchmark scores, however impressive, do not survive contact with users, and Meta has been burned before by models that looked strong in evaluation and disappointed in the real world. The company’s engineers are said to be confident; its competitors are watching the release date with more than usual interest.
For the industry, the Watermelon disclosures confirm what the training bills already suggested: the frontier of AI is now an infrastructure contest as much as a research contest. Meta has decided that catching OpenAI is worth ten times the compute. The question now is whether the model can justify the bill.
