Skip to content
NewsModel release

Meta's unreleased Watermelon model has matched GPT-5.5 on key benchmarks, Alexandr Wang tells employees

· by Pondero Newsdesk

The short version

Meta Superintelligence Chief Alexandr Wang told employees on July 2 that the company's in-training Watermelon model has reached parity with OpenAI's GPT-5.5 on internal benchmarks, using an order of magnitude more compute than its predecessor Avocado.

Meta's unreleased Watermelon model has matched GPT-5.5 on key benchmarks, Alexandr Wang tells employees

Meta Superintelligence Chief Alexandr Wang told employees during an internal town hall on July 2, 2026 that the company's next model, codenamed Watermelon, has reached parity with OpenAI's GPT-5.5 on closely watched AI benchmarks. The claim, reported first by Business Insider and corroborated by Benzinga and American Bazaar, comes from Wang directly and is grounded in internal benchmark results. Meta has not publicly confirmed which benchmarks were used.

What

Watermelon is Meta's next model after Avocado, the internal name for Muse Spark, which Meta released in April 2026. Wang said Watermelon "uses an order of magnitude more compute than Avocado," per the American Bazaar, citing a person familiar with the matter who attended the town hall. The model is still in training and has no confirmed public release date.

GPT-5.5, which OpenAI released in April 2026, remains the most broadly available OpenAI flagship for enterprise and consumer customers. OpenAI has since introduced GPT-5.6, but rolled out that model under restrictions requested by the Trump administration, making it effectively unavailable outside certain government contexts. Parity with GPT-5.5, if the internal benchmarks hold against independent evaluation, would place Watermelon at the front of the public-access competitive field.

Wang also posted on X around the same time that a Muse Spark update would ship soon, focused on coding and agentic capabilities. He noted the update was designed to close the gap with Anthropic's Claude Opus in coding tasks, per the American Bazaar report.

Wang oversees Meta Superintelligence Labs, an elite internal AI research team the company refers to as TBD. Meta hired Wang from Scale AI roughly nine months before the Muse Spark launch.

Why it matters

The Watermelon claim is notable because it is the first specific benchmark-parity assertion from a named Meta executive since the company restructured its AI org around Wang and scaled up compute investment. Prior coverage of Meta's AI push focused on Muse Spark as a competitive but not frontier-tier model. Watermelon, if the internal numbers survive external review, would represent a meaningful shift in that picture.

Two caveats apply. First, the specific benchmarks remain undisclosed. Benchmark selection shapes parity claims considerably. A model that matches GPT-5.5 on a narrow set of tasks may underperform on broader evaluations. Second, at the same town hall, CEO Mark Zuckerberg struck a noticeably more cautious note. Reuters reported that Zuckerberg told employees that "the trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected" and acknowledged that leadership had been overly optimistic about AI coding tools advancing quickly.

For AI tool operators choosing between frontier models, the practical effect is zero until Watermelon ships and independent benchmarks run. Operators who have built on Llama-series models should watch whether Watermelon follows the same open-weight path or ships as a closed API model.

Context

Meta launched Muse Spark in April and signaled it would open-source future versions of the Muse model family. Whether Watermelon follows that path has not been addressed publicly. Benzinga reported that Meta did not immediately respond to a request for comment on the benchmark claim. The Watermelon assertion is sourced to a named executive at an internal event and corroborated by multiple outlets, though it remains a self-reported internal result.

What to watch next

Two things are worth tracking. Whether Meta discloses the specific benchmarks behind Wang's claim and whether an independent evaluation, once Watermelon ships publicly, confirms the GPT-5.5 parity claim. Muse Spark's imminent update on coding and agentic tasks is the nearer-term milestone. Wang's X post did not give a date, but "soon" in that context likely means weeks, not quarters.

Sources