Meta's "Watermelon" AI Model Matches GPT-5.5 as Google Capping Forces Token Limits

Meta's "Watermelon" AI Model Matches GPT-5.5 as Google Capping Forces Token Limits

Meta Platforms is making rapid strides to close the frontier AI capability gap, even as it navigates severe infrastructure constraints and commercial friction with direct rivals. During an internal town hall on July 2, 2026, Alexandr Wang, Head of Meta Superintelligence Labs, announced to employees that Meta's next-generation flagship AI model—codenamed Watermelon—has successfully matched OpenAI’s flagship GPT-5.5 model on key internal benchmarks.

Watermelon represents a massive generational leap for Meta, utilizing an order of magnitude more compute than its predecessor, Avocado (the internal codename for Muse Spark, which was released in April 2026). This progress is part of Meta's broader strategy to eliminate its reliance on third-party AI models for internal operations.

The Google Gemini Rationing and the End of "Tokenmaxxing"

The urgency behind Watermelon's development has been heavily accelerated by a quiet infrastructure crisis. In March 2026, Google refused to sell Meta all the Gemini AI computing capacity it requested, citing an inability to meet Meta's exceptionally high demand. This capacity cap, which remains in place, severely disrupted Meta's internal workflows.

Historically, Meta had been renting Google's Gemini to handle major internal operations—including safety automation, scam detection, customer service, and developer coding workflows—because Llama models lagged in performance. The rationing forced Meta to abruptly end its high-spending "tokenmaxxing" era—where employees were encouraged to consume as many tokens as possible, burning through over 60 trillion tokens in a single 30-day period at an estimated annual cost of $50,000 per employee. Meta has since directed engineers to transition to its internal Muse Spark models and coding assistants to cut costs and mitigate the Gemini cap.

To permanently secure its independence, Meta has committed to a colossal physical infrastructure roadmap, planning up to $145 billion in capital expenditure for 2026 alone and a targeted $600 billion investment in the United States by 2028.

Verbatim Quotes

"Watermelon, our next model after Avocado, is currently in training... Watermelon uses an order of magnitude more compute than Avocado." — Alexandr Wang, Head of Meta Superintelligence Labs, quoted in Business Insider

"Google has refused to sell Meta all the Gemini AI computing capacity it wanted, telling the social media giant around March 2026 that it simply could not meet the demand¹... By one SemiAnalysis estimate, Meta's per-employee spending on AI tokens had reached roughly $50,000 annually at list prices, with employees burning through over 60 trillion tokens in a single 30-day period in early 2026." — The Times of India