Meta says its next AI matches GPT-5.5 performance

Meta's 'Watermelon' Signals a New Compute Arms Race

Meta is reportedly testing an internal model, codenamed "Watermelon," that the company claims performs on par with GPT-5.5 on its own benchmark suite. According to the reporting, the leap in capability comes largely from a roughly tenfold increase in computational power compared to Meta's previous generation of models, rather than solely from architectural breakthroughs.

Why This Matters

If accurate, this would mark a significant moment for Meta's AI ambitions. The company has positioned its Llama family of models as the standard-bearer for openly available large language models, competing against closed systems from OpenAI, Google, and Anthropic. A model that can match GPT-5.5-class performance would suggest Meta is closing the capability gap with the industry's most advanced closed labs — at least on internal metrics, which should always be read with some skepticism since companies design and select benchmarks that flatter their own systems.

The detail about a 10x compute increase is arguably the most important part of this story. It suggests Meta's gains here are being driven substantially by scaling — more GPUs, more training data, more energy — rather than efficiency innovations. That has implications beyond Meta's own roadmap: it reinforces the industry-wide pattern where frontier-level performance increasingly requires massive capital expenditure, favoring companies with the deepest pockets and largest data-center footprints.

The Open Source Angle

Meta has built its AI brand around releasing Llama models with permissive (if not fully open-source) licenses, arguing this democratizes access to advanced AI. Watermelon's existence raises an open question that will matter enormously to the open-source AI community: will Meta actually release this model, or a version of it, publicly?

A tenfold compute increase implies substantially higher training and likely inference costs, which could make Meta more cautious about giving away a frontier-level model for free, especially given competitive pressure and the costs already committed to infrastructure buildouts. If Meta does open-source a GPT-5.5-class model, it would be a landmark event, effectively handing the open community capabilities previously available only through paid APIs from closed labs. If it doesn't, or releases a distilled/smaller version, that would signal a shift in Meta's strategy toward treating its most capable models as proprietary assets.

What to Watch

Until independent, third-party benchmarking is available, claims of parity with GPT-5.5 should be treated as preliminary. Key signals to watch include whether Meta confirms Watermelon officially, what licensing terms accompany any release, and how the model performs on external evaluations rather than internal comparisons.