Meta's next frontier large language model, internally codenamed Watermelon, remains unreleased as of mid-October 2026 despite internal documents targeting a public launch that month. Alexandr Wang, Meta's chief AI officer, stated in a July town hall that the still-training model had reached parity with OpenAI's GPT-5.5 on unnamed internal benchmarks while using roughly 10 times the compute of the April-released Muse Spark predecessor. No external evaluations or official confirmation exist, and the model was absent from Meta Connect in September. Traders are weighing the October window against typical frontier training delays, recent Muse agent security fixes, Meta's massive chip and infrastructure spend, and competitive pressure from OpenAI and Anthropic. A credible announcement or weights release in the coming weeks would be the clearest catalyst.