The Secret Model
Business Insider reported on July 2, 2026 that Meta's Chief AI Officer Alexandr Wang told a closed briefing that Meta's model currently in training, internally codenamed 'Watermelon', matches GPT-5.5 performance on current evaluations. The headline detail: it uses an 'order of magnitude more' compute than Meta's previous frontier model.
What 'Order of Magnitude' Means
If Meta's previous frontier training run used approximately 100,000 H100 GPU-equivalents, Watermelon is training on roughly 1 million GPU-equivalents. Meta has the largest GPU fleet of any company outside of cloud providers, with an estimated 600,000 H100s in production and plans for a 1-million-GPU cluster announced earlier in 2026.
The Data Advantage
Watermelon's training data includes a significant proprietary dataset from Meta's social platforms — Facebook, Instagram, WhatsApp, and Threads — that no other lab can replicate. This gives Meta a unique edge in training data that encompasses global human communication patterns, visual content, and real-time interactions at a scale unmatched by any competitor.
GPT-5.5 Class Performance
The GPT-5.5 class performance claim, if accurate on independent benchmarks, would make Watermelon the highest-performing Meta model ever, surpassing Llama 4 on frontier reasoning and coding tasks. This positions Meta as a genuine frontier competitor — a significant escalation from its previous open-weight Llama strategy.
Timeline and Caveats
Wang's briefing occurred in a non-public setting. Business Insider's reporting is sourced to attendees. Meta has not made an official statement. The release timeline for Watermelon has not been confirmed. The model is still in training, meaning final capabilities may differ from current evaluations.
Sources: Business Insider (July 2, 2026); Meta AI infrastructure disclosures.
