Meta woke up? AI head: New model 'Watermelon' scores can match GPT-5.5

Meta Superintelligence Labs head Alexandr Wang revealed in an internal meeting that the new model "Watermelon" in training, using about 10x the computing power of Muse Spark, has matched the performance of OpenAI's GPT-5.5.
(Background: Goodbye Llama! Meta launches new multimodal AI model "Muse Spark," fully deployed on IG, FB, and smart glasses)
(Background: OpenAI's new model GPT-5.6 not allowed to launch: Trump administration demands phased release)

Table of Contents

Toggle

• From Open Source Retreat to Closed Source Sprint
• Ten Times the Computing Power, How Much Gap?
• The Finish Line Being Chased Has Been Moving All Along

According to a report from Business Insider citing two informed sources, Meta Superintelligence Labs head Alexandr Wang told employees in an internal all-hands meeting this week that the next-generation model codenamed "Watermelon" has matched the performance of OpenAI's flagship model GPT-5.5, released in April this year. Like Avocado launched in April, Watermelon is also a fruit codename commonly used internally at Meta.

Wang stated that Watermelon is the next model after Muse Spark (internal codename Avocado), which went live in April this year. It is still in training and uses an order of magnitude more computing resources than Muse Spark, which translates to nearly 10 times.

This is the first time in the past year that Meta has publicly compared one of its yet-to-be-released models to OpenAI's current flagship. He did not disclose which benchmarks were used to determine "matching"... Meta declined to comment on this, and OpenAI did not respond to requests for comment.

From Open Source Retreat to Closed Source Sprint

Meta's AI anxiety is nothing new. The company has been spending heavily on computing power, building data centers, and poaching talent. Over the past two years, it has been trying to carve a place among OpenAI, Google, and Anthropic, but the Llama series has always remained in a position of "decent scores, but still far from the top."

Last year, Meta spent $14.3 billion to acquire a 49% non-voting stake in Scale AI. This deal not only gave Meta access to Scale AI's data labeling capabilities, but also brought co-founder Alexandr Wang directly into the company as its first Chief AI Officer. After Wang took office, the original AI department was renamed Meta Superintelligence Labs. In addition to overseeing R&D, he also leads an elite research team codenamed TBD, as well as recently promoted hardware projects.

This talent arms race has also driven up the compensation levels Meta offers, raising the entire Silicon Valley AI industry's standards.

The first model Wang delivered after taking office was Muse Spark, released in April this year. This model has a fundamental difference from the previous Llama: Llama had open weights, while Muse Spark is completely closed source. Meta only stated that it "hopes future versions can be open source."

Its scores were not bad, but it still couldn't surpass OpenAI or Anthropic. After launch, the model was quickly integrated into flagship products like Instagram, Facebook, and smart glasses, allowing Wang's team's achievements to directly reach ordinary users for the first time, rather than just staying on benchmark leaderboards.

Ten Times the Computing Power, How Much Gap?

The numbers Wang gave at the internal meeting were straightforward: Watermelon uses an order of magnitude more training computing power than Muse Spark. Simply put, Meta is not fine-tuning this time; it is directly multiplying its bet by ten. This also echoes Wang's consistent logic: computing power determines the model's ceiling, scores are just the first checkpoint, and what truly decides the outcome is whether the model can later be put into the product line. He also sent similar signals on X:

First, Mark was clearly talking about the industry’s progress on agentic capabilities on the whole.

But, while we’re on the topic: Our next Muse Spark update is coming soon. Big improvements in coding and agentic capabilities to be more competitive with other leading models.… https://t.co/uTjx8sZM2A

— Alexandr Wang (@alexandr_wang) July 3, 2026

The post teases that Muse Spark will soon receive an update, focusing on coding ability and agentic tasks, with the goal of narrowing the gap with competitors.

A user asked Meta when it could deliver a coding model on par with Claude Opus. Wang replied "soon" and said people would like what they are "cooking." Whether this somewhat mysterious statement will be accepted by the market remains uncertain.

The Finish Line Being Chased Has Been Moving All Along

But somewhat awkwardly, GPT-5.5 is OpenAI's flagship model from April this year, but OpenAI already released the more powerful GPT-5.6 family at the end of June. However, due to the Trump administration's request, it is currently only available for preview to a small number of registered partners and has not been fully launched.

For OpenAI, this is not a technological bottleneck but a policy bottleneck. The model has already been built, but it is not yet allowed for full release. In other words, Watermelon has actually caught up with a model that OpenAI released over two months ago; the real ceiling has already moved up.

If Wang's statement is accurate, this is still the first signal Meta has been able to present after two years of heavy spending and poaching. Meta has already told investors that its spending on infrastructure such as chips and data centers this year will be revised up from the previous estimate of $115 billion to $135 billion to $125 billion to $145 billion; the signing bonuses for top researchers have even reached hundreds of millions of dollars per person.

This means that Zuckerberg's talent blitz may finally be seeing some returns. But scores are always the easiest goal to achieve. Whether Meta can actually bring Watermelon to the table and get developers to buy in is the threshold Meta has never been able to cross in the past two years.