Meta says its next AI matches GPT-5.5 performance

What's Being Claimed

Meta is reportedly telling insiders that its next-generation AI model, internally codenamed "Watermelon," performs on par with GPT-5.5 on internal benchmarks. According to the reporting, the model was trained using roughly ten times the computational power of Meta's previous-generation systems, suggesting the company is leaning heavily into brute-force scaling to close the gap with rivals like OpenAI.

Why This Matters

If accurate, this would mark a significant shift in Meta's AI trajectory. The company's Llama family has generally been positioned as a capable but not category-leading line of models, especially compared to frontier systems from OpenAI, Anthropic, and Google DeepMind. A model that matches GPT-5.5-level performance — even on Meta's own internal tests — would signal that Meta is no longer content playing catch-up and is willing to spend enormous compute budgets to compete at the very top of the field.

The 10x compute increase is the detail worth scrutinizing most. Scaling laws have historically produced real gains, but returns tend to diminish, and the industry has increasingly debated whether raw compute alone can keep delivering step-change improvements. If Meta needed an order-of-magnitude jump in resources just to reach parity with a competitor's model, it raises questions about efficiency and cost-effectiveness relative to labs that may be achieving similar results with more targeted architectural or data improvements.

The Benchmark Caveat

It's important to stress that these are described as internal benchmarks — not independent, third-party evaluations. Internal testing is useful for engineering teams but is inherently self-reported and can be shaped by which benchmarks are chosen, how they're weighted, and how comparably they're run against a competitor's model. Historically, claims of matching or beating a rival model on internal tests haven't always held up once external researchers or public leaderboards get access. Until Watermelon (or whatever it's eventually named at release) is tested independently — on things like reasoning benchmarks, coding evaluations, or real-world user preference tests — this should be treated as a preliminary signal rather than confirmed parity.

Context and What to Watch

Meta has been investing aggressively in AI infrastructure, including massive data center buildouts and chip purchases, so a jump in compute allocation for a flagship model fits its broader strategy. The real test will come when Watermelon — or its eventual public release — is benchmarked against GPT-5.5 and other frontier models under neutral conditions. Watch for details on release timing, whether Meta open-sources any version of the model as it has with past Llama releases, and how it performs on widely-used public benchmarks like MMLU, GPQA, or coding-specific tests once available.