Meta claims its AI model matched OpenAI's—but left out the proof

Meta's Alexandr Wang told company staff that the firm's in-development AI model, codenamed Watermelon, has reached parity with OpenAI's GPT-5.5 on key performance metrics. The claim, reported internally and later surfaced to media outlets, offered little concrete evidence—no benchmark tables, no model documentation, and no public validation.

A claim without the crucial detail#

Wang named the rival company, identified the model being trained, and delivered a verdict of competitive parity. What he omitted, however, was the single most important element: which benchmarks demonstrated this equivalence. When Business Insider reporters asked the two attendees who disclosed the claim, neither could specify which tests Wang had referenced. Meta declined to comment further, and OpenAI did not respond.

The absence of a benchmark table marks a sharp departure from how rival labs typically announce such milestones. OpenAI published detailed results when launching GPT-5.5, and Anthropic releases evaluation data with each Claude Opus version. A public benchmark table transforms a boast into something the research community can examine, debate, and validate. An internal score shared only with staff amounts to a morale-boosting number rather than a replicable finding.

Context matters—and counts against the narrative#

Taking Wang's statement at face value actually undermines the achievement he described. Watermelon allegedly requires roughly ten times the computational resources of Meta's previous Muse Spark model to match GPT-5.5's performance. Yet by the time Wang made this comparison, OpenAI had already advanced to a limited preview of GPT-5.6—the successor that made GPT-5.5 obsolete. Meta's effort to match a model from weeks prior, while consuming vastly more compute, looks less like progress and more like playing catch-up.

The timing matters too. Meta faces earnings season this quarter, and staff morale sits raw following nearly 8,000 layoffs and 7,000 reassignments earlier in the year. An upbeat message from the superintelligence chief, untethered to public proof, reads as reassurance more than breakthrough.

What actually landed: admission, not achievement#

Wang followed the town hall with a social media post promising substantial gains for an updated Muse Spark in coding and autonomous agent capabilities—again, with ambition outpacing hard numbers. The only verified statement to emerge from Meta's internal gathering that week came from CEO Mark Zuckerberg's simultaneous admission to staff: that the company's AI investments "haven't come to fruition yet." Until Watermelon ships with published benchmarks that hold up under scrutiny, Zuckerberg's candid assessment remains the story worth believing.