Meta's next-generation AI model, code-named "Watermelon," has caught up with OpenAI's flagship GPT-5.5 on key benchmarks, according to AI chief Alexandr Wang. The claim, reported by Business Insider, offers a rare bright spot for a company that has spent heavily to close the gap with rivals. Yet the same internal town hall that produced the model update also exposed deep frustration at the top: CEO Mark Zuckerberg admitted that Meta's broader push into AI agents is moving more slowly than he had planned.
The dual messages capture a company at a crossroads. For more on the competitive pressures shaping the industry, follow our breaking AI news coverage as the frontier-model race intensifies.
Watermelon: An Order-of-Magnitude Compute Bet
Speaking to staff at an internal town hall on Thursday, July 2, Wang said Watermelon has reached parity with OpenAI's top model, GPT-5.5, citing benchmarks he did not publicly specify. "Watermelon, our next model after Avocado, is currently in training," Wang said, according to Business Insider. "Watermelon uses an order of magnitude more compute than Avocado."
"Avocado" is the internal code name for Muse Spark, the first model in Meta's new frontier lineup, which shipped in April. Muse Spark posted solid benchmark scores but failed to match offerings from OpenAI or Anthropic, leaving Meta perceived as a step behind. The decision to pour roughly ten times the compute into its successor signals how aggressively the company is betting that raw scale can close the performance gap.
The stakes are unusually high. Meta plans to spend up to $145 billion on AI infrastructure this year, a sizable chunk of the more than $700 billion that Big Tech collectively is pouring into the field.
Zuckerberg Concedes Agents Are Behind Schedule
Wang's upbeat model update clashed with a more sober assessment from Zuckerberg, whose remarks were captured in an audio recording obtained by Reuters. The CEO acknowledged that the AI agents Meta has reorganized its business around have not advanced as fast as expected.
"The trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected," Zuckerberg told employees, according to Reuters. He added that the bets on the company's new structure "haven't come to fruition yet" and that the corporate restructuring "didn't go as clean as it could have."
The admission carries real weight because of how much Meta has wagered. Over the past year, Zuckerberg placed Wang in charge of the AI division, rebranded it as Meta Superintelligence Labs, and offered nine-figure compensation packages to lure top researchers away from OpenAI and Google. In May, the company laid off roughly ten percent of its global workforce and moved about 7,000 employees into AI teams, aiming to fund infrastructure and squeeze efficiency from AI-powered workflows.
A Rosier Picture From the AI Chief
At the same town hall, Wang struck a markedly different tone from his boss. On the social platform X, he moved into damage-control mode, arguing that Zuckerberg had been speaking about the progress of the entire AI industry rather than Meta's efforts specifically.
Wang pointed to tangible near-term releases. A Muse Spark update with major improvements to coding and agentic capabilities is coming soon, he said, along with a coding model that he claimed is on par with Anthropic's Claude Opus "pretty soon." He suggested users would appreciate what the team has been "cooking."
The Longer Arc Behind the Code Names
The internal model code names trace Meta's difficult journey. In March, The New York Times reported that Meta had delayed an earlier model, also called Avocado, after internal tests showed it fell short of Google, OpenAI, and Anthropic in logical reasoning, programming, and writing. Leadership at one point even discussed temporarily licensing Google's Gemini, though no deal materialized. Muse Spark eventually shipped in April, and Watermelon was positioned as the next leap forward.
Whether Watermelon can deliver on Wang's benchmark claims in independent evaluations remains to be seen. Meta has historically released models without open weights, a departure from its earlier open-source posture, which means outside researchers may have limited ability to verify the figures before the model reaches users.
What Comes Next
Zuckerberg told employees he expects more tangible results within the next three to six months. Meanwhile, Meta is exploring adjacent revenue streams to help finance its AI ambitions. Bloomberg reported that the company is building a cloud business to sell excess AI compute capacity to outside customers, mirroring a playbook already pursued by other infrastructure-rich firms.
Separately, Chief Technology Officer Andrew Bosworth addressed the company's controversial mouse-tracking software, which records employee activity to generate AI training data. An internal review found that no employee data had entered AI training, Bosworth said. If the program restarts, it will run on an opt-in basis.
For Meta, the coming months will test whether bigger models and bigger spending can finally translate into products that match the company's ambitions.
Stay Ahead of AI
For the latest on frontier AI models, company strategy, and the competitive landscape, visit AI Buzz Wire.
Read more AI news →
