Key Highlights –
- Meta has pushed the release of its next-generation AI model, codenamed Avocado, from March to at least May 2026, citing performance that falls short of rival models
- Sources familiar with the matter told Reuters that Avocado’s current performance sits between Google’s Gemini 2.5 and Gemini 3 far from the competitive benchmark Meta was aiming for
- In the interim, Meta is reportedly considering licensing Google’s Gemini to power its own AI products which is a significant admission given the scale of investment the company has poured into building its own frontier model
Meta had positioned the first quarter of 2026 as its moment to reassert itself in the AI race. That moment has now been pushed back by at least two months, and the circumstances around the delay reveal that the company is under considerably more strain than its public communications have suggested.
The New York Times, citing three people with knowledge of the matter, reported Thursday that Avocado, Meta’s next-generation text AI model developed inside the newly formed TBD Lab, will not launch until May at the earliest. A Meta spokesperson offered a measured response: the next model “will be good, but more importantly, show the rapid trajectory we’re on.” That framing is notable for what it doesn’t say. It is not a denial of the delay, and it is not a denial of the performance gap.
What the Performance Gap Actually Means
The core issue, per sources cited by Reuters, is that Avocado currently benchmarks between Google’s Gemini 2.5 and Gemini 3, the latter of which launched in November 2025. Releasing a model that sits below a competitor’s four-month-old release would be a difficult position for a company that has publicly committed $115 to $135 billion in AI capital spending for 2026 alone.
To be clear, Avocado reportedly does outperform Meta’s own previous models. That matters for internal progress tracking, but it is not the bar Meta set when it restructured its AI organisation, brought in Alexandr Wang from Scale AI at a reported cost of $14.3 billion, and added marquee names like former GitHub CEO Nat Friedman and Shengjia Zhao, a co-creator of ChatGPT. Those hires were sold to investors as the foundation of a lab with, in Zuckerberg’s own words from October, “the highest talent density in the industry.” A delay grounded in below-par benchmarks is an early and uncomfortable test of that claim.
The Gemini Licensing Question and Why It’s the Most Telling Detail
The detail that deserves the most attention in this story is not the delay itself. It is the reported consideration of temporarily licensing Google’s Gemini to power Meta’s AI products while Avocado catches up.
Meta has spent years building the case that open-source development through its Llama model family was both philosophically principled and strategically sound. Llama’s wide adoption gave Meta credibility in the developer community and genuine ecosystem traction. Avocado was supposed to be the next chapter and there were already signals it might launch as a closed, proprietary model, abandoning the open-source posture that Llama represented.
Now the company is reportedly weighing whether to temporarily run a competitor’s model inside its own products. For users of Meta AI across WhatsApp, Instagram, and Facebook, that would mean responses generated by Google infrastructure, branded as Meta AI, with no visible indication of the switch. Whether Meta would disclose that arrangement to users is a question the company hasn’t addressed publicly, but it may not have a choice everywhere. Under the EU AI Act’s transparency provisions, users interacting with AI systems have the right to know when AI is involved in a response, and member states are already moving toward enforcement. A silent model swap behind a Meta-branded interface could put the company on a collision course with European regulators before Avocado even launches.
A Pattern Worth Noting
Avocado has now slipped three times. The original target was sometime in 2025, which moved to early 2026, which has now moved again to May at the earliest. Each delay has been accompanied by reassurances of strong internal progress – and that pattern does not make the current situation catastrophic. Model development timelines routinely shift, and no serious analyst expects a single quarter to define a company’s long-term AI position. But it does make each reassurance incrementally harder to take at face value.
The broader context is also relevant. Meta’s stock fell 2.55% on Thursday’s news. Competitor models have continued to move. OpenAI is advancing GPT-5, Google’s Gemini 3 already has a four-month head start on Avocado, and Anthropic’s Claude 4 has drawn strong reception in the enterprise space. The window for Meta to announce a credible frontier model (rather than just a better version of its own previous work) is narrowing with each passing quarter.
