Meta Superintelligence Labs head Alexandr Wang told employees in an internal all-hands meeting that the company’s next model, code-named “Watermelon,” has caught up with OpenAI’s GPT-5.5 on benchmarks. According to his remarks, the model is still in training, but it is being built with roughly 10 times the compute used for Muse Spark, which Meta released in April.
The report came from Business Insider, citing two people familiar with the matter. “Watermelon,” like “Avocado” before it, follows Meta’s internal practice of using fruit codenames for models. Wang did not specify which benchmarks were used to support the comparison. Meta declined to comment, and OpenAI did not respond to a request for comment.
Meta shifts from open-weight roots to closed-model push
The statement stands out because Meta has spent the past year pouring money into compute, data centers, and recruiting as it tries to close the gap with OpenAI, Google, and Anthropic. Its Llama family has remained competitive on paper, but it has not been viewed as leading the field.
Last year, Meta bought a 49% non-voting stake in Scale AI for $14.3 billion. The deal gave Meta access to Scale AI’s data-labeling capacity and also brought co-founder Alexandr Wang into the company as its first Chief AI Officer. After he arrived, Meta renamed its AI division Meta Superintelligence Labs. He now oversees research, a team code-named TBD, and the company’s recent hardware push.
Wang’s first major model at Meta was Muse Spark, released in April. Unlike Llama, which uses open weights, Muse Spark was fully closed-source, with Meta only saying that future versions might become open source. After launch, the model was quickly integrated into Instagram, Facebook, and Meta’s smart glasses, putting the work in front of mainstream users rather than keeping it inside benchmark rankings.
A bigger compute bet, but the target has moved
Wang’s internal description was blunt: Watermelon is being trained with an order of magnitude more compute than Muse Spark. That suggests Meta is not making a minor tuning pass. It is scaling the bet sharply. Wang has hinted at the same direction on X, where he said Muse Spark would soon get upgrades focused on coding ability and agent tasks in an effort to narrow the gap with rivals.
When users asked when Meta might ship a coding model on the level of Claude Opus, Wang replied, “soon,” adding that people would like what the company was “cooking.” For now, though, the only concrete claim is the internal benchmark comparison. Meta has not disclosed launch timing or detailed capability data for Watermelon.
OpenAI has already advanced beyond GPT-5.5
The comparison also comes with an important caveat. GPT-5.5 was OpenAI’s flagship model in April, but the report says OpenAI introduced the stronger GPT-5.6 family in late June. That release has not been broadly rolled out and is limited to a small group of pre-approved partners under U.S. government requirements.
If Wang’s claim is accurate, Watermelon is matching OpenAI’s earlier flagship rather than its newest available generation. Even so, it would still mark one of the clearest signs yet that Meta’s spending spree is producing visible results. Meta has already told investors that its expected 2026 spending on chips, data centers, and related infrastructure would rise from $115 billion-$135 billion to $125 billion-$145 billion. The company has also reportedly offered top researchers signing packages worth hundreds of millions of dollars per person.
Watermelon has not been launched. The benchmark claim may be an early indicator, but the harder test for Meta is whether the model can make it into products and win adoption from developers.