Meta’s chief AI officer Alexandr Wang told employees last week that the company’s next flagship model — internally codenamed “Watermelon” — has matched OpenAI’s GPT-5.5 on key performance benchmarks. The model is still in training, which means Meta is claiming performance parity before a single line of it has shipped to the public. The disclosure came during an internal company town hall and was first reported by Benzinga and multiple AI outlets.
What Is Meta’s Watermelon Model?
Watermelon is the direct successor to Meta’s “Avocado” — the internal name for Muse Spark, released in April 2026. Where Avocado was already a meaningful step up, Watermelon uses “an order of magnitude more compute,” according to Wang. That phrasing alone suggests Meta is treating this as its true frontier push, not a mid-cycle refresh. The model is still in active training and carries no confirmed release date as of today.
Benchmarks — But Which Ones?
The catch: Wang did not name the specific benchmarks Watermelon cleared during the town hall. Meta declined to elaborate, and OpenAI did not respond to requests for comment. That makes the claim difficult to independently verify right now. OpenAI’s GPT-5.5 set a high bar on reasoning and coding tasks when it launched earlier this year, and OpenAI has since pushed out GPT-5.6 Sol, Terra, and Luna, though that rollout remains limited to roughly 20 government and partner organizations. If Watermelon is matching GPT-5.5 today, the gap to GPT-5.6 is still real and measurable.
The AI Race Is Compressing Fast
The significance of this leak is what it signals about timing: the window in which any single model dominates is now measured in weeks, not quarters. Microsoft’s MAI models are competing at the enterprise tier. Grok 4.5 is already in private beta at SpaceX and Tesla, built on a 1.5T-parameter V9 foundation. And Chinese open-source models like GLM-5.2 are closing the gap from the open-weights side. The market is converging at the frontier faster than anyone predicted at the start of 2026.
Our Take
Unverified benchmark claims from internal town halls are a well-worn move in the AI industry — they warm up the analyst community and rattle competitors before a model actually ships. Read Wang’s statement as a signal of intent, not a confirmed result. That said, the fact that Meta chose GPT-5.5 as its reference point — not some abstract internal target — suggests genuine confidence. Watermelon is coming, the compute scale is serious, and the race at the frontier is about to get more crowded.
What Is Meta Watermelon AI?
Meta Watermelon is the internal codename for Meta’s next-generation AI flagship model, currently in training as of July 2026. Meta’s chief AI officer Alexandr Wang disclosed at an internal town hall that the model has matched OpenAI’s GPT-5.5 on key benchmarks, though the specific evaluations were not named publicly. Watermelon uses significantly more compute than Meta’s previous frontier model, Avocado (Muse Spark), and is expected to launch later in 2026.
