Meta's 'Watermelon' AI Model Catches GPT-5.5 — But the Race Has Already Moved On

The problem? Wang did not disclose which benchmarks were used, making the claim unverifiable . Without knowing whether Watermelon matched GPT-5.5 on coding benchmarks (where Meta's Muse Spark scored only 59.0 on Terminal-Bench versus GPT-5.4's 75.1), PhD-level reasoning tests like GPQA Diamond (where Muse Spark scored 89.5%), or general language understanding, the comparison is incomplete .

Why GPT-5.5 Is Already a Moving Target

OpenAI released GPT-5.5 on April 23, 2026, where it quickly topped the Artificial Analysis Intelligence Index . But by the time Wang made his statement, OpenAI had already launched GPT-5.6 on June 26, 2026 — a three-model family (Sol, Terra, Luna) with the flagship Sol setting a new state of the art on Terminal-Bench 2.1 at 88.8% (and 91.9% in its Ultra configuration) .

This means Watermelon is targeting a model that is no longer the frontier. OpenAI's cadence — GPT-5.4 on March 5, GPT-5.5 on April 23, GPT-5.6 on June 26 — suggests a roughly six-week iteration cycle . By the time Watermelon ships, OpenAI may already be at GPT-5.7 or beyond.

Meta's Track Record: Avocado's Troubles

Watermelon's bullish claim comes against the backdrop of Meta's struggles with its predecessor. Avocado — Meta's first major model from its restructured Superintelligence Lab — was delayed from March 2026 to at least May 2026 after internal tests showed it lagged behind Google's Gemini 3.0, OpenAI's GPT-5.4, and Anthropic's latest models on reasoning, coding, and writing tasks . One report indicated Meta even considered temporarily licensing Google's Gemini as a fallback .

Muse Spark (Avocado's public name) eventually launched but placed fourth globally behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 on the Artificial Analysis Intelligence Index, scoring 52 out of 100 . It excelled on medical and scientific benchmarks (42.8% on HealthBench Hard, best in class) but struggled on coding (59.0 on Terminal-Bench) and abstract reasoning (42.5 on ARC AGI 2) .

The Massive Spending Question

Zuckerberg has bet heavily on AI infrastructure. Meta's Superintelligence Lab reportedly spent $14.3 billion on its first major model (Muse Spark) . Wang's comment that Watermelon uses 'an order of magnitude more compute' than Avocado suggests Meta is willing to throw enormous resources at closing the gap .

However, on the same day reports emerged of Wang's bullish claim, Mark Zuckerberg reportedly expressed concern that AI progress was slowing, creating a mixed internal message about Meta's trajectory . The tension between Wang's optimism and Zuckerberg's caution highlights the uncertainty around whether raw compute scaling alone will close the gap with OpenAI's faster iteration pace.

GPT-5.6 Adds Another Wrinkle

OpenAI's GPT-5.6 Sol is currently available only as a 'limited preview' to select partners under US government directives . This means its full capabilities may not be rapidly commercialized — but it still sets a higher performance bar than GPT-5.5 . Sol's Terminal-Bench 2.1 score of 88.8% (91.9% Ultra) versus GPT-5.5's 83.4% represents a meaningful capability jump .

Bottom Line: A Rearward-Looking Target

Wang's claim that Watermelon matches GPT-5.5 is likely true on some benchmarks, but it represents a rearward-looking target. By the time Watermelon ships, OpenAI may already be at GPT-5.7 or beyond, and Meta's track record of execution delays (Avocado) combined with enormous capital spending creates genuine uncertainty about whether raw compute scaling alone will close the gap with OpenAI's pace of iteration. The real question isn't whether Watermelon matches GPT-5.5 — it's whether Meta can sustain the infrastructure investment needed to keep pace with a competitor that releases a new frontier model every six weeks.