Meta’s latest AI push is no longer just catching up. The company’s superintelligence chief, Alexandr Wang, told employees that Watermelon, the model Meta is currently training, has already matched OpenAI’s GPT-5.5.
Watermelon is still in training and runs on roughly 10 times the compute of its predecessor, Muse Spark, which launched in April. That extra scale appears to be paying off. On top of the base model, Wang teased an incoming Muse Spark update with, quote, big coding and agentic gains, with the release expected to hit both Meta AI and the company’s new API.
The timing matters. CEO Mark Zuckerberg made headlines at the same town hall by saying agent progress hasn’t really accelerated in the way that we expected. Wang later clarified that Zuckerberg was talking about the industry’s overall agent progress, not Meta’s. He followed with a promise: expect an Opus-level coding model pretty soon.
Why this matters
A GPT-5.5-level model would be a meaningful jump for Meta. Its current Muse Spark model already sat comfortably beneath the field even at launch, and investors have been watching whether Meta’s reported $145 billion AI spend would eventually translate into competitive frontier systems.
If Wang’s claims hold, the answer may finally be yes. The catch is that the frontier itself keeps moving. Anthropic’s Mythos and Fable models are already showing what the next step up looks like, and OpenAI’s 5.6 models are expected this week. That means Watermelon may close one gap only to face another.
Still, the combination of a stronger base model, a focused coding upgrade, and Zuckerberg’s public admission suggests Meta is serious about competing at the top again. The next few months will show whether that seriousness turns into sustained leadership.