Meta committed $135 billion to AI this year, brought Scale AI's co-founder Alexandr Wang aboard as its Chief AI Officer, and poured money into a company that might reshape how frontier models are built. None of it was enough. According to reporting by Reuters and the New York Times, Meta's next-generation model — codenamed "Avocado" — was quietly delayed from a March launch to at least May 2026 after internal benchmarks revealed its performance still falls short of competitors. The delay is more than a scheduling slip. It's a window into the structural difficulty of catching up in a race where the leaders are accelerating.
What Avocado Was Supposed to Be
Meta began working on Avocado in the second half of 2025, positioning it as the flagship of its rebuilt AI research organization. After the company's LLaMA-series models established Meta as a credible open-source AI player, Avocado was meant to be something different: a proprietary, performance-first model designed to compete directly with Google's Gemini 3, OpenAI's GPT-5.4, and Anthropic's Claude family on the closed-model frontier benchmarks that actually move the industry.
The company had originally targeted a Q1 2026 release, according to a Reuters report from January. Internal delivery of early models was confirmed by Meta's CTO. Zuckerberg himself set expectations publicly, telling investors during the company's January earnings call that Meta's next model would be "good" and would "show the rapid trajectory we're on." What it hasn't shown, at least not yet, is the ability to beat — or even fully match — what competitors already have in production.
The Performance Gap
According to people familiar with the matter who spoke to the New York Times, Avocado's performance currently falls somewhere between Google's Gemini 2.5 and Gemini 3 — meaning it is more capable than a model Google released in mid-2025, but not yet competitive with Gemini 3, which Google launched in November. That gap matters because Google, OpenAI, and Anthropic have all continued shipping improvements in the interim.
The failure modes are specific: Avocado underperformed on internal benchmarks for reasoning, coding, and writing — three of the most commercially and technically consequential capabilities in modern AI models. Reasoning and coding benchmarks have become the primary competitive battleground in 2025 and 2026. GPT-5.4, which OpenAI shipped on March 5, scored above human-level performance on OSWorld-Verified computer use tasks. Anthropic's Claude Code has become the benchmark tool for agentic software development. Avocado is not, by current internal assessments, in the same conversation.
Meta's spokesperson offered a statement that was carefully worded in exactly the way a company speaks when the news is bad but the strategy is not: "As we've said publicly, our next model will be good, but more importantly, show the rapid trajectory we're on, and then we'll steadily push the frontier over the course of the year as we continue to release new models. We're excited for people to see what we've been cooking very soon." The language is notable for emphasizing trajectory over current state — a signal that May's release, when it arrives, will be framed as part of a series rather than a singular competitive statement.
The $135 Billion Question
The delay lands at a moment when Meta's AI spending has become almost incomprehensible in scale. The company confirmed in its January 2026 earnings call that it plans to spend between $115 billion and $135 billion in capital expenditures this year, up from $72 billion in 2025 — itself a record. The primary driver is AI infrastructure: data center buildout, chip acquisition, and talent.
That spending spree includes some audacious bets. In June 2025, Meta finalized a $14.3 billion investment in Scale AI, the data labeling and AI evaluation firm founded by Alexandr Wang, and brought Wang aboard as Chief AI Officer. Scale AI's core competency is exactly the kind of rigorous model evaluation that benchmarks like the ones Avocado reportedly failed are designed to probe. The irony of hiring the world's foremost AI evaluator and still shipping a model that fails internal benchmarks is not lost on observers.
Meta has also been building its own silicon. In March 2026, the company unveiled plans for its next generation of in-house AI chips — continuing a strategy of reducing dependence on Nvidia GPUs for both training and inference. But custom silicon takes years to design, validate, and deploy at scale. For this model cycle, Meta is still training on external hardware, and the training outcomes have not met expectations.
The Gemini Licensing Option
Perhaps the most remarkable detail in the reporting is this: according to the New York Times, leaders of Meta's AI division have discussed the possibility of temporarily licensing Gemini from Google to power Meta's AI products while Avocado continues development. No final decisions have been made. But the fact that the conversation is happening at all is a significant signal.
For context: Meta's consumer AI products — including the Meta AI assistant embedded across WhatsApp, Instagram, Facebook, and the Ray-Ban smart glasses — are currently powered by Meta's own Llama-series models. If Meta were to license Gemini even as a stopgap, it would be an admission that its own models are not competitive enough to power its own flagship consumer products. That's a strategic concession of enormous significance for a company that has publicly committed to AI self-sufficiency and "superintelligence" as a corporate north star.
The Gemini licensing discussion also reveals something about competitive dynamics: Google, which builds and deploys Gemini against Meta's own AI assistant in the consumer market, would effectively be powering its competitor's products. This kind of arrangement is not without precedent in tech — Microsoft built its business on IBM's distribution for years — but in the current AI race, it reads as a sign of how far ahead Google has pulled on model capability.
Why Throwing Money Doesn't Buy a Frontier Model
The deeper story in the Avocado delay is what it reveals about the structural economics of frontier AI development. Meta is arguably the largest AI spender in the world on an absolute basis, yet it is trailing organizations like Anthropic — which has burned through more than $10 billion on training and inference against only $5 billion in lifetime GAAP revenue — on model capability.
Part of the explanation is organizational. Google and Anthropic have concentrated their research talent on a small number of flagship model families, with clear internal hierarchies of purpose. Meta's AI organization has historically been more distributed: separate teams working on LLaMA, on multimodal systems, on the AI assistant, on internal infrastructure. That distribution is a strength for open-source breadth but a weakness for the focused push required to build a single frontier-competitive model.
Part of the explanation is also technical. Training a frontier model is not primarily a function of spending — it is a function of data quality, architectural choices, and an evaluation pipeline that reliably distinguishes genuine capability improvements from benchmark artifacts. Alexandr Wang's appointment was meant to address the evaluation side. The fact that Avocado failed its internal benchmarks suggests the evaluation pipeline is working as intended; what's lagging is the training outcome itself.
And part of the explanation is simply that the frontier has moved faster than Meta anticipated. When Avocado's training run was planned, Google's Gemini 3 did not yet exist. OpenAI's GPT-5.4 computer-use capabilities did not yet exist. The models that were credibly competitive in early 2025 are not competitive against what Google and OpenAI shipped in late 2025 and early 2026. The bar moved, and Avocado was trained to clear a lower one.
What Comes Next
A May or June 2026 launch window gives Meta roughly two to three months to close the gap — or, more likely, to set expectations carefully about what Avocado represents in a longer model roadmap. Zuckerberg's "rapid trajectory" framing suggests that strategy: present Avocado as the first in a series of improving releases rather than a single competitive statement.
The open-source angle may also figure prominently. Meta's LLaMA series has been enormously influential in the research community, and a strong LLaMA-4 or Avocado open release could generate goodwill and ecosystem momentum even if the closed-model benchmarks are not best-in-class. That's a different kind of value proposition than "beat GPT-5" — but it may be the more achievable one in the near term.
For now, though, the AI race's most expensive competitor is running behind schedule. In a field where a single quarter of delay can mean irrelevance — where Google can ship Gemini 3 in November and make the Avocado training run obsolete — that is not a small problem. Meta has the capital to keep competing. What remains to be seen is whether capital, alone, is the variable that decides who wins.
