Avocado was supposed to show that Meta’s superintelligence push was paying off. But now it’s late, and Gemini may fill the gap
Chesnot/Getty Images
Meta $META’s latest AI fruit is ripening on company time. The New York Times reported Thursday that Meta has pushed back Avocado — the text model meant to headline CEO Mark Zuckerberg’s latest and greatest AI reset — from a planned March debut to at least May after internal testing showed it trailing the leading systems from Google $GOOGL, OpenAI, and Anthropic on, essentially, everything: reasoning, coding, and writing.
That’s another ugly moment for a company that has spent lavishly to convince investors, employees, and Silicon Valley that it could buy its way back into the front rank of AI. In January, Zuckerberg told investors that Meta’s first new models would be “good” and would show the “rapid trajectory” the company was on. That same month, the company said it expects 2026 capital expenditures of $115 billion to $135 billion, with the increase driven in part by Meta Superintelligence Labs. But writing giant checks and building giant data centers still doesn’t guarantee a place in the top tier of the model race.
The Times, citing unnamed sources familiar with the matter, reported that Avocado currently lands somewhere between Google’s Gemini 2.5 and Gemini 3. That’s a rough place to end up after months and months of buildup because it suggests no real progress on the one thing this project was supposed to do: make AI people talk about the company as a peer again.
Building a frontier model is hard. Shipping one that can actually stand next to the best systems from Google, OpenAI, and Anthropic is harder. Meta’s problem is that it has made the stakes of that challenge impossible to ignore. Meta needs a model that changes the conversation around the company. Avocado, at least for now, seems to be doing the opposite.
This isn’t Meta’s first awkward moment in the lab. Llama 4 was delayed after failing to meet Meta’s expectations on reasoning and math benchmarks, even after Zuckerberg said — in January 2025 — that he expected it to “become the leading state-of-the-art model” that year. The model wasn’t particularly well received. In May, Meta delayed its flagship Behemoth model because engineers were struggling to produce significant enough improvements. By late June, Zuckerberg had reorganized the company’s AI work under Meta Superintelligence Labs. Now, Meta’s leaders are already thinking past Avocado. The next model, the Times reported, is being called Watermelon.
Leaders inside Meta’s AI division have discussed temporarily licensing Google’s Gemini to power some of the company’s AI products while Avocado catches up, The Times reports. When the company spending like it wants to own the next computing platform is reportedly discussing renting Google’s brain for the interim, the catch-up story starts sounding even more expensive than its bills already are.
Meta has already tried the expensive reset. This was supposed to be a quarter when the company started showing that all the spending bills, all the futuristic promises, all the talent drama, and all the superintelligence rhetoric were producing something more persuasive than ambition. Instead, the picture now looks uncomfortably familiar. And Avocado looks like another reminder that throwing billions of dollars at a problem doesn’t make the benchmarks blink.
Meta may still close the AI gap. Rich companies get more than one swing, and Zuckerberg clearly intends to keep paying for the at-bats. The company has vast distribution and huge infrastructure plans. But the public scorecard right now shows a company that has spent heavily, hired aggressively, reorganized repeatedly, and still has to explain why its next big model is late — and why one of the options on the table was leaning on the rival it’s eventually supposed to beat.
Investors have been willing to indulge Meta’s AI appetite because the core business keeps printing money. The ad machine remains a monster, and that cash flow has helped make the company’s AI splurge feel, if not exactly calm, then at least defensible. When Meta reported fourth-quarter results in January, shareholders rewarded the company even as it laid out a spending plan that would have looked deranged almost anywhere else. Zuckerberg called 2026 a big year for delivering personal superintelligence. That sort of line works a lot better when the next major model shows up on time and looks like a genuine leap.
But the frontier-model race is different. This is the prestige contest. This is where Google, OpenAI, and Anthropic get to look like the grown-ups, and where Meta has spent the past year trying to prove it shouldn’t be sitting at the kids' table anymore. A delayed flagship model doesn’t exactly help make that case.
For now, Avocado looks like another reminder that frontier AI has been a humbling business for Meta. The company has money, compute, distribution, and one of the most determined CEOs in tech. What Meta still doesn’t have is the clean, undeniable launch that makes all of that spending look like foresight instead of impatience.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
