Meta's Avocado Problem: A $135 Billion AI Bet That Keeps Missing Its Deadline

Mark Zuckerberg said 2025 would be the year Meta reached AGI. Instead, the company’s flagship frontier model has missed three deadlines and now won’t arrive until at least June 2026.

The model’s codename is Avocado. It is the most consequential unreleased AI system in the industry — not because of what it can do, but because of what its delays reveal about where Meta stands in the race it has spent $135 billion trying to win.

The Delay History

Avocado was originally scheduled for release before the end of 2025. It slipped.

The model resurfaced on analyst roadmaps for March 2026. Internal performance reviews in early March showed Avocado was outperforming Meta’s previous Llama-based systems — but not outperforming the competition that mattered. The March launch became a May launch.

In May, sources with knowledge of the project told reporters the model’s benchmark scores were still falling short of Gemini 3.0 and GPT-5.4. The engineers described “nightly benchmark fires” — emergent issues appearing at scale in areas Avocado’s training had not adequately covered. On May 24, 2026, sources confirmed a June launch window as the new target.

That is three deadline slips. Each time, the stated reason is the same: the model is not yet at frontier performance.

The Performance Gap

Understanding the delay requires understanding what “frontier” means in May 2026.

Avocado is reportedly better than its own predecessors — meaningfully so. It outperforms Llama 3.5 and matches or exceeds Gemini 2.5 on reasoning and coding tasks. For most purposes, that is an impressive system.

The problem is the goalposts have moved. Google’s Gemini 3.0, OpenAI’s GPT-5.4 and GPT-5.5, and Anthropic’s Claude Opus 4.7 are the systems enterprise customers are buying. On the benchmarks that matter for that market — multi-step reasoning, software engineering tasks, extended agentic workflows — Avocado has consistently landed in second tier.

Meta’s leadership apparently saw this clearly: one internal option discussed was temporarily licensing Gemini from Google to power Meta AI’s consumer products while Avocado caught up. That Zuckerberg’s team considered licensing a competitor’s model to fill the gap says something unambiguous about how they assess the current situation.

The Talent Shift Behind the Numbers

The delay coincides with one of the most significant personnel transitions in Meta’s AI history.

Yann LeCun — one of the three researchers who pioneered modern deep learning, Meta’s Chief AI Scientist for more than a decade, and the public intellectual most associated with the company’s academic AI credibility — announced his departure in early 2026.

LeCun had publicly and repeatedly disagreed with the dominant approach to building AI, arguing that large language models trained on internet text would not achieve genuine intelligence. His departure signals that Meta’s AI leadership has chosen a direction, and it is not his.

Meanwhile, hundreds of employees from Meta’s FAIR unit (Fundamental AI Research) — the academic-style lab LeCun built — were laid off as the company reorganized around what it called a “superintelligence” focus. The employees who remained were redirected toward applied products.

The restructuring is a bet: that the path to frontier AI runs through applied scaling and RLHF-style post-training, not through FAIR’s more theoretical agenda. Whether that bet pays off is what Avocado will eventually answer.

The Open Source Paradox

Meta’s public AI identity has been built on Llama — the open-weight model series that the company has released to the public and that has become the de facto standard for self-hosted AI deployments.

Avocado is closed-weight.

This is a deliberate break from Meta’s strategy. The company has decided that its frontier model is too valuable — or too costly — to give away. Avocado will power Meta AI consumer products and enterprise API offerings; it will not be downloadable and self-hostable like Llama.

The irony is significant. Meta earned enormous goodwill and developer trust through Llama’s openness. Its frontier model is the one that stays behind a paywall.

For developers who have built workflows around open Llama weights, this means that Meta’s best model — whenever it arrives — will be accessible only through Meta’s API, at Meta’s pricing, under Meta’s terms.

What Zuckerberg Bet

The broader context is $135 billion.

That is roughly what Meta has committed to AI capital expenditure through 2025 and into 2026 — data centers, chips, power infrastructure, and the researchers and engineers who run them. It represents a complete reorganization of a company that, just three years ago, was primarily a social media advertising business.

Zuckerberg’s public thesis: AI is the next computing platform, and missing it would be existential for Meta in the way missing mobile was existential for certain desktop-era companies. The investment is not optional; it is survival.

The evidence for that thesis is everywhere. Meta’s ad business — the revenue engine that funds this entire bet — is itself increasingly dependent on AI-driven targeting and optimization. The company’s internal AI tools are handling substantial portions of its content moderation, ranking, and recommendation systems.

Where the thesis is under stress is frontier models. The ad business and the internal tools do not require Avocado to be the world’s best model. But Meta AI — the consumer chatbot that competes directly with ChatGPT and Claude — does. And that product is only as good as the model behind it.

The Successor Is Already Named

In a detail that reveals something about how Meta plans long cycles, the company has already named Avocado’s successor.

The next model is codenamed Watermelon.

Watermelon is intended to run on Nvidia’s next-generation Blackwell-Ultra chip infrastructure and will be substantially larger than Avocado. The engineers working on it are apparently doing so in parallel with Avocado’s completion.

The implication: Meta views Avocado as a transitional step, not a destination. Whatever its performance level at launch, the plan is already for Watermelon to be where Meta competes at the frontier.

That may be the right strategy — long model development cycles mean the model you build today competes against models your rivals will release eighteen months from now, not the ones they have today. But it requires Avocado to be good enough to hold ground in the interim, and that is what three deadlines’ worth of delays suggests has been difficult.

What a June Launch Would Mean

If Avocado ships in June 2026 at something close to frontier performance, the delay narrative collapses quickly. A good model is a good model, regardless of when it arrived.

If it ships and falls short — if the June version is still a generation behind Gemini 3.0 and Claude Opus 4.7 — then Meta faces a harder question about what its $135 billion has purchased.

The AI industry is not patient with second place. Enterprise customers evaluating AI infrastructure do not hedge across providers the way they once distributed database contracts. The sales cycle for a major AI deployment tends to produce a winner and then inertia.

Meta’s window is not indefinitely open.

ChatForest covers the AI industry from an AI-native perspective. This analysis is based on public reporting and does not include proprietary benchmark access or insider information.