- Meta superintelligence chief Alexandr Wang reportedly told employees that Watermelon, the model currently in training, has matched OpenAI’s GPT-5.5.
- Watermelon runs on roughly 10 times the compute of its predecessor Muse Spark, which launched in April.
- Wang clarified on X that Zuckerberg’s town-hall remark about slow agent progress referred to the industry as a whole, and teased an “Opus-level” coding model “pretty soon.”
- A Muse Spark update with “big coding and agentic gains” is planned for both Meta AI and the company’s new API.
What Happened
Meta superintelligence chief Alexandr Wang reportedly told employees that Watermelon, the AI model the company is currently training, has matched OpenAI’s GPT-5.5, according to a July 6, 2026 report from The Rundown. Wang said the model is still in training and runs on roughly 10 times the compute of its predecessor, Muse Spark, which launched in April.
Why It Matters
The claim lands days after CEO Mark Zuckerberg acknowledged at the same internal town hall that agent progress “hasn’t really accelerated in the way that we expected” — a concession widely read as an admission that Meta’s AI reorganization has underdelivered. Wang clarified on X that Zuckerberg meant the industry’s agent progress as a whole, not Meta’s specifically.
A GPT-5.5-class model would be a meaningful jump for Meta: Muse Spark posted solid benchmark scores at launch but sat clearly below the leading systems from OpenAI and Anthropic. If Wang’s claim holds through release, Meta’s reported $145 billion AI spend would finally be producing a model the rest of the field has to take seriously.
Technical Details
Per The Rundown’s account, Watermelon is mid-training at roughly 10x Muse Spark’s compute budget, and Wang’s team is also preparing a nearer-term Muse Spark update with “big coding and agentic gains,” to ship on both Meta AI and the company’s new API. In a reply on X, Wang told followers to expect an “Opus-level” coding model — a reference to Anthropic’s Claude Opus tier — “pretty soon.” The comparison point matters: all claims so far are internal, made mid-training, and not yet backed by public benchmarks or third-party evaluation.
Who’s Affected
The announcement targets Meta’s own researchers and investors first — a morale and confidence signal after Zuckerberg’s candid town-hall assessment — and developers second, who would get the coding-focused Muse Spark update through Meta AI and the new API. Competitively, the claim puts OpenAI, Anthropic, and Google on notice that Meta believes it has closed the gap to at least the GPT-5.5 tier, though the frontier is not standing still: Anthropic’s Mythos and Fable models mark the next step up, and OpenAI’s 5.6 models are expected imminently.
What’s Next
Watermelon remains in training with no announced release date, so the checkable milestones are the Muse Spark update Wang promised for Meta AI and the API, and the teased coding model. The limitation to keep in view: matching GPT-5.5 mid-training is a moving target, since the models Watermelon will actually launch against are the ones shipping months from now, not today’s. Meta’s claim will be measurable only when the model reaches public benchmarks.
