The New York Times reported on March 12, 2026 that Meta has delayed its flagship AI model, internally codenamed “Avocado,” after the model failed to meet performance benchmarks in internal testing. It is the second consecutive model disappointment for a company that has staked its near-term future — and spent more than any organization in history — on AI leadership.

What ‘Avocado’ Was Supposed to Be

“Avocado” was designed to match GPT-5.2 and Claude Opus 4.6 — a fully multimodal flagship capable of competing on reasoning, code generation, and visual understanding. It was supposed to validate Meta’s $60B infrastructure investment and give the three billion users of WhatsApp, Instagram, Facebook, and Messenger a genuinely competitive AI assistant.

Instead, internal test results fell short. The model has been sent back for additional work. No revised timeline has been announced publicly.

Why Llama 4 Underperformed — and the Pattern

Llama 4 launched in April 2025 to underwhelming benchmarks. Analysts cited training data quality issues, over-optimization on benchmark tasks, and talent gaps relative to OpenAI and Anthropic. At the time, the shortfall was treated as a one-time stumble. Avocado’s failure suggests the problem is structural.

The Open-Source Trap: When You Can’t Hide Benchmarks

Meta’s open-source strategy — releasing model weights publicly — creates a specific vulnerability: performance ceilings are immediately visible to the entire research community. Closed frontier labs can quietly iterate on failures. Meta cannot.

Chinese open-source models — notably Qwen 3.5 and GLM-5 — are now benchmarking at or above Llama 4. Meta’s open-source strategy risks positioning Meta AI as a second-tier offering in its own distribution channels.

What 3 Billion Users Are Actually Running

Meta AI is deployed across WhatsApp, Instagram, Facebook, and Messenger — over three billion users. The gap between distribution and model quality is becoming a reputational and product problem.

Zuckerberg called 2025 “the year of AI.” Meta’s $60B infrastructure spend was the largest single-year AI capital expenditure in history, backed by approximately $200B in cash reserves.

Model Performance Comparison

Can Meta Course-Correct?

Meta has resources that no other company can match: $200B in cash, the world’s largest social graph, and distribution to three billion users. The question is whether the company can translate infrastructure spend into model quality before the gap with frontier models becomes too visible to ignore.

The “Year of Efficiency” (2023) showed Meta can execute decisive operational pivots. What’s less clear is whether the capability problems are tractable with more compute — or whether they require different talent, different architecture, or a different relationship with safety and alignment tuning.