Meta delays Avocado AI model after internal tests lag rivals

Meta just hit the brakes on its next big AI release.

According to The New York Times, Meta has delayed its next-generation AI model, internally codenamed "Avocado." Originally scheduled to launch in mid-March 2026, the model will not ship until May at the earliest, citing three people familiar with the matter.

The culprit? Internal benchmarks showed Avocado simply couldn't keep pace with the competition.

Falling behind Google and OpenAI

Inside Meta's testing labs, Avocado reportedly fell short of top-tier models from Google, OpenAI, and Anthropic across logical reasoning, coding, and writing.

To be fair, Avocado wasn't a total failure. It beat Meta's previous model and outpaced Google's Gemini 2.5. But it couldn't stand up to Gemini 3.0.

The shortfall was serious enough that Meta's leadership actually discussed temporarily licensing Google's Gemini to bridge the gap, though no final decision was made.

Meta officially stated that updates are coming "very soon" and noted that more models are on the schedule for this year. The company's roadmap also includes a future flagship codenamed "Watermelon" and an image and video generator dubbed "Mango."

A $14B shakeup and an open-source pivot

Meta used to dominate mindshare with its open Llama models, but that momentum stalled after Llama 4 failed to keep up.

In response, CEO Mark Zuckerberg poured tens of billions into AI, including a massive $14.3 billion investment in Scale AI. Scale AI CEO Alexandr Wang now runs Meta's frontier lab, known as "TBD Lab," with the mission of building superintelligent systems.

With new leadership and mounting performance pressure, reports also indicate that Meta may be backing away from the open-source philosophy that defined its early AI strategy.

Why it matters

We think Avocado's delay reveals just how punishing the frontier AI race has become, even for a company with Meta's balance sheet.

Spending $14.3 billion and reshuffling leadership under Alexandr Wang hasn't instantly solved the core intelligence gap. Falling behind Gemini 3.0 in reasoning and coding proves that throwing money at data and compute cannot guarantee you leapfrog frontier labs.

Even more critical is what this means for the broader ecosystem. If Meta actually pivots away from open source because its models are struggling to compete, developers and startups who relied on Llama as a free alternative to proprietary APIs could soon be left out in the cold. For now, anyone waiting on Meta's next leap to upgrade their workflows should expect to wait at least until May, if not longer.