Meta delayed the Avocado AI model because it failed internal performance tests, underperforming rivals Google, OpenAI, and Anthropic on reasoning, coding, and writing tasks. The model, originally planned for a March 2026 launch, has now been pushed to at least May.
This is the second delay. Meta initially targeted a 2025 release, shifted to early 2026, and has now slipped again. Internal tests show Avocado performs somewhere between Google’s Gemini 2.5 and Gemini 3, better than older models but not yet competitive at the frontier.
What is the Meta Avocado AI Model?
Meta Avocado is a next-generation large language model (LLM) being developed inside Meta Superintelligence Labs, a new internal AI division Meta created after restructuring its research teams in 2025. The model is designed to rival OpenAI’s GPT-5, Google’s Gemini 3, and Anthropic’s Claude in coding, reasoning, and writing.
Avocado is being built alongside a companion model called Mango, a multimodal image and video generation system. Both are the first flagship models from Meta’s new AI leadership team.
The project is led by Alexandr Wang, founder of Scale AI, whom Meta hired in June 2025 for $14.3 billion, one of the largest talent acquisitions in tech history. Wang heads Meta Superintelligence Labs, which also includes former GitHub CEO Nat Friedman and ChatGPT co-creator Shengjia Zhao.
Why is the Meta Avocado AI Model Delayed?
Avocado Failed Internal Performance Benchmarks
According to a New York Times report citing three people familiar with the matter, Avocado underperformed on internal tests for reasoning, coding, and writing. It was not competitive with the latest offerings from Google, OpenAI, and Anthropic when measured against current benchmarks.
The model beat Meta’s older systems and outperformed Google’s Gemini 2.5, but it could not match Gemini 3, which Google released in early 2026 and has received strong reviews from developers.
Post-Training Problems, Not Pre-Training
Meta’s team completed Avocado’s pre-training phase late in 2025. The bottleneck is in post-training, the fine-tuning stage where a model is refined for real-world tasks. The team began post-training in January 2026, initially targeting a mid-March release. That target has now been missed.
Post-training is where reasoning, instruction-following, and safety alignment are developed. Issues at this stage are common, but costly when competitors are shipping updates quickly.
Internal Strategy Confusion
Beyond the technical problems, Meta is also dealing with a strategic debate: should Avocado be open source or closed?
Meta built its reputation on open-source AI through the Llama model series. But Llama 4, released in April 2025, disappointed developers and failed to generate the momentum Meta expected. Since then, internal discussions have shifted toward a proprietary model approach, a significant departure from Meta’s public AI identity.
That debate is unresolved. And an unresolved strategy adds friction to development timelines.
Meta Is Considering Licensing Google’s Gemini
In a striking development first reported by the New York Times, Meta’s AI leaders have discussed temporarily licensing Google’s Gemini technology to power some of Meta’s consumer AI products while Avocado catches up. No decision has been made, but the fact that it was discussed internally signals how far behind the model currently sits.
How Does Avocado Compare to Competitors?
| Model | Company | Status |
| GPT-5 | OpenAI | Released and advancing |
| Gemini 3 | Released, strong developer reviews | |
| Claude (latest) | Anthropic | Released |
| Avocado | Meta | Delayed to May 2026 at earliest |
Avocado was expected to compete at the frontier when launched. Right now, based on internal reports, it sits between Gemini 2.5 and Gemini 3, which means it would launch already behind Google’s current best model.
How Much is Meta Spending on AI?
Meta’s 2026 capital expenditure plan is $115 billion to $135 billion, nearly double its 2025 spend of $72.2 billion. Over the past two years, Meta has deployed approximately $107 billion in AI infrastructure spending.
Despite that investment, Avocado has not yet matched the performance of competitors. As analysts at KeyBanc noted, Meta entered 2025 as an AI leader and now faces “more questions around investment levels and ROI.”
Meta’s stock slipped roughly 1% in Friday premarket trading after the NYT report broke.
What Did Meta Say About the Delay?
Meta has not officially confirmed the delay. A spokesperson told Reuters: “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.”
A separate spokesperson told CNBC: “Our model training efforts are going according to plan and have had no meaningful timing changes.”
Both statements push back on the delay narrative, but multiple sources cited by the New York Times, Reuters, and CNBC contradict the public messaging.
What Happens Next for Meta Avocado?
- Launch target: May 2026 at the earliest, though the timeline has shifted multiple times
- Open source vs. closed: Still undecided internally
- Gemini licensing: Under discussion but no decision made
- Mango (image/video model): Also in development; no updated timeline disclosed
The delay matters beyond one product launch. Meta has spent aggressively and restructured its entire AI division around delivering a competitive frontier model. Avocado is that model. A further slip, or a launch that underperforms GPT-5 and Gemini 3, would raise serious questions about whether Meta’s AI strategy is working, and whether its $14.3 billion bet on new leadership is paying off.
Key Facts: Meta Avocado AI Model Delay
- Original launch target: Q1 2025 → Early 2026 → Now May 2026 at earliest.
- Reason for delay: Failed internal benchmarks on reasoning, coding, and writing.
- Performance level: Between Gemini 2.5 and Gemini 3, not yet frontier.
- Team: Meta Superintelligence Labs, led by Alexandr Wang (Scale AI founder).
- Investment: $14.3 billion in Scale AI; $115-$135 billion 2026 capex planned.
- Gemini licensing: Under internal discussion as a stopgap.
- Sources: New York Times, Reuters, CNBC, all citing independent insiders.
