Meta's Watermelon AI model has caught up to GPT-5.5, Alexandr Wang says

July 3, 2026 - 05:20

Meta's ambitious push into advanced artificial intelligence appears to be paying off. According to Alexandr Wang, the company's head of superintelligence, its upcoming "Watermelon" model has now caught up to OpenAI's GPT-5.5 on critical performance benchmarks. The statement, made during a private industry briefing, suggests Meta is closing the gap with leading frontier labs faster than many analysts expected.

Watermelon, which has been in development for over a year under a tight secrecy wrapper, is designed to handle complex reasoning, long-context tasks, and multimodal inputs. Wang noted that internal evaluations show the model matching or exceeding GPT-5.5 on standard tests for logical deduction, code generation, and factual accuracy. "We are seeing parity in areas that matter most for real-world deployment," he said, without providing exact scores.

The achievement marks a significant milestone for Meta, which has historically trailed behind OpenAI and Google in the race for cutting-edge AI. The company has been investing heavily in compute infrastructure and talent, poaching researchers from rival firms. Watermelon is expected to power a new generation of Meta products, including enhanced virtual assistants and content moderation tools.

However, Wang cautioned that benchmark results do not tell the whole story. "Benchmarks are a useful snapshot, but they don't capture safety, alignment, or cost efficiency," he added. Meta plans to release a technical paper detailing Watermelon's architecture and limitations later this year. The model is still undergoing red-teaming and safety evaluations before any public rollout.