Zuckerberg Admits Meta's AI Agent Push Has Stalled Even as 'Watermelon' Model Claims GPT-5.5 Parity

In a candid admission that sent Meta's share price down 4.9 percent to close at $582.90, CEO Mark Zuckerberg told an internal company town hall on July 2 that Meta's work on AI agents had not 'accelerated in the way we expected' over at least the last four months. A recording of the meeting was reviewed by Reuters. Zuckerberg acknowledged that the company's structural bets — executed through approximately 8,000 layoffs representing roughly 10 percent of Meta's corporate workforce and the reassignment of a further 7,000 employees to AI-focused teams, including one called Agent Transformation — had not yet yielded the velocity in product delivery that executives had anticipated. He told staff he had made mistakes in the workforce transition and would 'almost certainly make more,' while projecting that meaningful returns from Meta's AI spending would arrive within three to six months.

Minutes later, in the same meeting, Meta's AI chief Alexandr Wang delivered a starkly different message. Wang, who heads the company's newly created Superintelligence Labs division, told employees that Meta's next foundation model, codenamed Watermelon, had caught up with OpenAI's GPT-5.5 on unspecified benchmarks, according to a Business Insider report. Wang said Watermelon uses an order of magnitude more compute than Avocado — Meta's April 2026 model, also referred to internally as Muse Spark — which itself had failed to match the frontier models from OpenAI or Anthropic on its initial release. The juxtaposition of Zuckerberg's candour about product delays and Wang's benchmark optimism in back-to-back addresses to the same audience created what multiple analysts described as a sharply divided internal narrative at precisely the wrong moment.

The structural problem analysts have identified is that even if Watermelon's benchmark claims are validated — and neither Meta nor OpenAI has publicly confirmed them, and no specific benchmark data was cited in the original report — the competitive landscape has already moved. OpenAI released GPT-5.6 in late June, meaning that a model reportedly matching GPT-5.5 would be chasing a target that was itself superseded before Watermelon reaches general availability. As analysts at FourWeekMBA noted, matching a model OpenAI has already superseded, while deploying an order of magnitude more compute than the prior generation, is less a capability breakthrough than a cost-structure problem dressed as a milestone. The dominant competitive question in AI has quietly shifted from who builds the best model to who builds the best model per dollar of compute — and on that metric, Meta's trajectory this week raised questions rather than answered them.

The capex context amplifies the stakes. Meta has raised its 2026 AI capital expenditure forecast to a range of $125 billion to $145 billion, making it one of the largest single-year technology infrastructure commitments in corporate history. Zuckerberg's three-to-six-month payoff window is now a self-imposed internal deadline, one that employees and investors will measure the company against through the back half of 2026. The appointment of Wang — formerly CEO of data-labelling firm Scale AI — to lead Meta Superintelligence Labs in 2025 was presented as an aggressive pivot to close the gap with Google and OpenAI. The July 2 town hall suggested that pivot has been more costly and slower to yield results than its architects had publicly projected.

The broader implications extend beyond Meta's organisational challenges. Zuckerberg has been among the most prominent executives arguing that autonomous agents represent the next major computing platform shift, and Meta has spent accordingly on that thesis. When the person authorising that level of capital tells staff the timeline has slipped, it becomes a data point about the industry-wide agent thesis, not just Meta's org chart. Rival labs, investors, and enterprise customers building their own agent strategies will be watching closely whether Meta's three-to-six-month window produces the acceleration Zuckerberg promised, or whether the gap between raw foundation model capability and genuinely useful deployed agents proves wider and more stubborn than the most optimistic forecasts assumed.