Meta Superintelligence Labs chief Alexandr Wang told employees at an internal town hall that Watermelon, the company's next flagship model, has caught up with OpenAI's GPT-5.5 on benchmarks. Business Insider first reported the claim, citing people who were in the room. The model is still training.
That is the headline. The caveats are longer than the headline.
What Wang actually said, and didn't
Wang described Watermelon as the successor to Avocado, the internal codename for Muse Spark, the model family Meta shipped in April. Watermelon reportedly burns an order of magnitude more compute than its predecessor, roughly tenfold, drawing on Meta's Prometheus cluster in Ohio. So the story here is not a clever architectural breakthrough. Meta poured a lot more compute into a bigger run and, per Wang, reached parity.
He did not name the benchmarks. He did not give a ship date. Meta declined to comment, OpenAI didn't respond, and OpenAI has already partly moved past GPT-5.5 with a late-June limited preview of GPT-5.6. Parity with a model your rival shipped in April and is already lapping is a strange flex, though reaching it at all is more than Muse Spark managed.
The numbers Muse Spark posted
Worth grounding this in what Meta's current model actually scores. Independent firm Artificial Analysis placed Muse Spark at 52 on its Intelligence Index. Respectable, top five globally, a real recovery from the widely panned Llama 4. But GPT-5.5 scored 59 on the same index, Gemini 3.1 Pro hit 57, and Claude Opus 4.6 landed at 53. Coding was where Meta's gap showed most.
Which is the whole point of the second announcement.
The Muse Spark update
Wang went public on X on July 3, saying the next update to the current Muse Spark is coming soon, with what he called big improvements in coding and agentic capabilities. He framed it as clarifying Zuckerberg's town hall remark that agent development hadn't accelerated the way leadership expected. Asked by a user when Meta would match Anthropic's Claude Opus on coding, Wang said the upcoming update is built to narrow exactly that gap. It rolls out through Meta AI and a new API.
"Big improvements in coding and agentic capabilities to be more competitive with other leading models." That's the pitch, more or less verbatim, and it reads like a man trying to get ahead of his own CEO's bad news.
Because the timing is awkward. Zuckerberg, per Reuters, told the same town hall that Meta's AI investments haven't paid off as fast as he anticipated. He runs cautious, Wang runs bullish, and both statements came out of the same room on the same day. Meta now projects $125 billion to $145 billion in 2026 infrastructure spending, up from an earlier $115 to $135 billion range. That gap between the two men is more informative than either quote alone.
Should anyone believe it yet
No independent evaluation exists. No published benchmark table, no model card, no third-party testing. One executive, one internal room, one round of unnamed benchmarks. Take it as a directional signal, not a datapoint. Under neural scaling laws, a tenfold compute bump produces real gains but not a tenfold leap, so "caught up on some benchmarks" and "beats Opus on real coding work" are very different sentences.
The thing to watch is whether Meta publishes a benchmark table when Watermelon ships. If it does and the numbers survive contact with independent evals, the frontier stops being a two-horse OpenAI-and-Anthropic race, which changes procurement math for anyone locked into either vendor. The Muse Spark coding update arrives first, via Meta AI and the new API, with no firm date beyond "soon." Watermelon remains in training with no announced release.
