Meta delayed its Avocado AI model to May 2026 after failing internal benchmarks
The system lags behind OpenAI, Google, and Anthropic in reasoning and coding
Chief AI Officer Alexandr Wang is leading a shift toward proprietary, paid models
Meta Platforms has delayed the release of its next foundational AI model, code-named Avocado, pushing its expected launch to at least May after internal tests showed the system lagging behind rival models on key capabilities such as reasoning, coding and writing tasks, The New York Times reported.
The delay complicates what was expected to be a breakthrough year for the company’s AI push and tempers expectations that Meta would soon match the performance of leading large language models from competitors such as Google, OpenAI and Anthropic.
2 March 2026
Get the latest issue of Outlook Business
Avocado’s Performance Issues
Internal evaluations reportedly found that Avocado performs better than Meta’s earlier model, Llama 4, but still trails the most recent systems released by rival AI developers.
The performance gap has led executives to consider interim measures, including the possibility of licensing an external AI model to power some products while development of Avocado continues.
The setback comes amid a broader restructuring of Meta’s AI strategy. Following the underwhelming reception of Llama 4, the company invested $14.3 billion in Scale AI and elevated its founder Alexandr Wang to help lead the company’s AI initiatives.
Meta’s TBD Lab
A new internal group known as TBD Lab has been tasked with developing several next-generation models, including Avocado as well as an image- and video-focused model called Mango. The team is also planning a larger follow-up model known internally as Watermelon. However, the effort has reportedly faced internal turnover and strategic disagreements with senior executives such as Chris Cox and Andrew Bosworth over priorities and monetisation strategy.
The situation highlights how even companies with vast financial and computing resources struggle to quickly match the performance of cutting-edge AI systems. Achieving strong performance across multiple skills, including reasoning, coding and natural language generation, remains technically challenging.
The delay also reflects Meta’s ongoing debate over whether its future models should be open source or more tightly controlled. While earlier Llama models were widely released to the developer community, the company is reportedly leaning toward more controlled releases to better manage safety and commercial considerations.
