Meta’s unreleased Watermelon model, the successor to Muse Spark from its Superintelligence Labs, remains the central driver of trader sentiment on any Humanity’s Last Exam debut. Internally reported in July 2026 as matching OpenAI’s GPT-5.5 on unnamed benchmarks after training with roughly 10× the compute of its predecessor, the model has not yet shipped or undergone public HLE evaluation as of late September 2026. Traders are watching Meta’s heavy infrastructure spend—projected at $125–145 billion this year—and rumored October release windows, which could position Watermelon as the next frontier entry on the demanding graduate-level benchmark where current leaders like Anthropic’s Claude models score in the low-to-mid 60% range. Any verified HLE results or launch confirmation would directly shift odds.