Meta's Watermelon, the internal codename for its next frontier large language model after Muse Spark, remains in late-stage training with reported use of roughly 10× more compute than its predecessor and internal claims of matching OpenAI's GPT-5.5 on unnamed benchmarks. As of early October 2026, no public release, weights, or external evaluations have occurred despite a targeted October window and a recent Mark Zuckerberg social-media tease linking the model to the Muse agent platform. Humanity's Last Exam (HLE) serves as one of the most demanding public benchmarks still separating frontier systems, with top models scoring below 60% even with tools. Trader sentiment hinges on whether Meta ships Watermelon soon enough to publish first-mover results on HLE or similar hard evaluations ahead of competitors, amid heavy infrastructure spending and ongoing competition from OpenAI and Anthropic.