Meta and its upcoming model, codenamed Watermelon, have reportedly caught up with OpenAI’s flagship model, GPT-5.5. Wang announced this at an internal meeting with employees, citing closely monitored benchmark tests. He did not specify which ones.
Meta desperately needs to show progress in AI. The company has poured billions of dollars into chips, data centers, and poaching experts from competitors, but it has yet to convince developers or customers that its models belong in the top tier. If Wang’s claim is true, it would be the strongest evidence yet that Meta’s efforts are beginning to pay off.
Wang described the status of the new model to employees. “Watermelon, our next model after Avocado, is currently being trained,” he said, according to a person familiar with the meeting. He added that Watermelon will use roughly ten times more computing power than Avocado. That is the internal designation for Muse Spark, the first model in this series that Meta released in April.
Muse Spark performed reasonably well in tests, but it failed to outperform either OpenAI or Anthropic. Watermelon’s tenfold increase in computing power looks like an attempt to bridge that gap through brute force. And Meta can afford it. The company told investors this year that it would spend between $125 billion and $145 billion on chips, data centers, and other equipment. It had originally expected to spend less, around $115 billion to $135 billion.
Wang wrote on X that he is preparing an update to the current Muse Spark. He promised significant improvements in coding and so-called agentic capabilities, meaning the system’s ability to complete multi-step tasks independently. The goal is to close the gap with competitors. When one user asked when Meta would have a coding model comparable to Anthropic’s Claude Opus, Wang replied, “pretty soon.” He added that people would like what the company was working on.
First, Mark was clearly talking about the industry’s progress on agentic capabilities on the whole.
— Alexandr Wang (@alexandr_wang) July 3, 2026
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
What happened at the same employee meeting is particularly interesting. While Wang painted a rosy picture of the future, Mark Zuckerberg admitted the opposite. According to a recording obtained by Reuters, he said that the development of AI agents had not accelerated over the past four months or so as he had expected. In his words, the bets on the new structure have not yet paid off.
Zuckerberg acknowledged that the company’s reorganization, which included layoffs, had not gone as smoothly as it could have. According to him, management got the timing wrong. Meta laid off roughly one-tenth of its workforce this year and reassigned around seven thousand people to teams focused on artificial intelligence. This caused dissatisfaction among employees.
When the company’s restructuring was being prepared in January and February, managers were “very optimistic,” according to Zuckerberg, about tools such as Anthropic’s Claude Code. However, that optimism has yet to translate into Meta’s own progress. Zuckerberg added that he expects to see greater benefits from AI investments over the next three to six months.
The disagreement between the two men was visible enough that Wang addressed it directly on X. He wrote that Zuckerberg had been talking about the progress of the industry as a whole, not specifically about Meta’s efforts. He then added his promise of the upcoming Muse Spark update. The market quickly formed its own opinion. Meta’s shares fell by almost five percent on Friday after the reports were published. The AI chief’s internal optimism has therefore not yet convinced investors.
It is worth recalling what the competition looks like. OpenAI released GPT-5.5 in April this year. Last June, it then introduced its most powerful model yet, GPT-5.6, but due to requirements from the US government, it is not yet publicly available. With Watermelon, Meta is therefore aiming at a target that continues to move further ahead.
Muse Spark, which Meta is building on, arrived in April as the first model from the Meta Superintelligence Labs division. Zuckerberg created it after reportedly becoming dissatisfied with how far the Llama models were falling behind ChatGPT and Claude. He put Wang, who previously led Scale AI, in charge. Meta invested $14.3 billion in the company and acquired a 49 percent stake. Wang now leads Meta’s elite research team, known as TBD, and also oversees other areas, including the development of proprietary hardware.
Sources: businessinsider.com and timesofindia.indiatimes.com
