Meta's unreleased Watermelon model, the successor to Muse Spark under Meta Superintelligence Labs, drives trader focus through internal claims of benchmark parity with OpenAI's GPT-5.5 despite remaining in training. Alexandr Wang highlighted an order-of-magnitude increase in compute over prior models during a July 2026 town hall, positioning it as a scaled frontier effort amid competition from Anthropic's Claude Opus series and OpenAI's GPT-6 line. No public release or independent evaluation has occurred by late September 2026, with Meta Connect passing without announcement and internal targets pointing to October. Humanity's Last Exam, the 2,500-question multimodal benchmark emphasizing graduate-level reasoning across math, physics, and other domains, serves as a key differentiator where current Meta models trail leaders. Potential catalysts include any October debut, safety reviews, or HLE-specific results that could shift competitive positioning.