Meta has delayed its next flagship AI model, internally codenamed Avocado, from a planned March release to May 2026. The reason, according to reports, is straightforward: internal testing showed the model underperforming against current offerings from Google, OpenAI, and Anthropic, particularly on reasoning, coding, and writing tasks.
The delay is notable for a company that has positioned itself as the leader of open-source AI. Meta's LLaMA model family underpins hundreds of startups and third-party tools, and a weak next-generation release would undermine both that ecosystem and Meta's credibility in the model race.
The timing compounds Meta's challenges. The company is simultaneously planning layoffs that could affect 20% of its workforce, with Reality Labs taking the deepest cuts, in part to redirect spending toward AI infrastructure. So Meta is cutting staff to fund AI development while also acknowledging that its AI development is not yet keeping pace with competitors.
For the broader AI tool market, Meta's struggles highlight just how expensive and difficult it is to stay competitive at the frontier. Google has Gemini, OpenAI has the GPT-5.4 family, and Anthropic has Claude. Each of these companies is spending tens of billions annually on compute.