Meta Says In-Training Watermelon Model Now Matches GPT-5.5 on Coding

Alexandr Wang, Meta's head of AI, told employees during a company meeting that the model currently in training under the codename Watermelon has matched OpenAI's flagship GPT-5.5 in coding capability. It's the first concrete claim of this kind since Meta released Muse Spark in April 2026, its first major model built under Wang's leadership.

What Wang Said

According to a Business Insider report also picked up by other outlets, Wang told employees directly that Watermelon, the successor to Avocado, is currently in training and consumes an order of magnitude more compute than Avocado did. Avocado was the internal codename for the model that launched publicly as Muse Spark.

Wang also announced a faster-arriving update to Muse Spark itself, focused on improving coding and agentic capabilities. It's expected to reach Meta AI users and developers using the new API before Watermelon fully rolls out.

Catching Up to Rivals

Meta has spent months trying to close the gap with OpenAI, Google, and Anthropic in the race for the best language models, even after Mark Zuckerberg's company poured tens of billions of dollars into superintelligence efforts and paid a sum in the billions to poach Wang from Scale AI. April's Muse Spark was seen as proof the new team could ship a product, but it lagged the leaders on coding benchmarks.

Watermelon is meant to change that. Matching GPT-5.5 on coding tasks, if confirmed by independent testing, would mean Meta now has a third full-fledged frontier model alongside OpenAI and Anthropic, with Google DeepMind as a fourth contender. According to people who attended the same meeting, Zuckerberg was noticeably more cautious in his assessments than Wang.

What It Means for Developers

The Muse Spark update is set to arrive in Meta AI and the new API, opening the door to using the model in developer tools, much as Claude Code and OpenAI's Codex already do. Analysts note that a cheaper or openly-weighted version of a Watermelon-class model could lower the cost of coding assistants for companies and reduce reliance on a single vendor.

For Polish development teams and software companies currently relying mainly on Claude Code, GitHub Copilot, or Cursor, this would mean another, potentially cheaper alternative in a market where subscription prices keep climbing alongside token consumption.

What's Next

Meta has not given a specific release date for either Watermelon or the promised Muse Spark update, saying only "coming soon." Until the model shows up in independent benchmarks such as SWE-bench or LiveCodeBench, claims of matching GPT-5.5 remain Meta's own internal assessment, based on leaks to US media rather than publicly verified results.

Sources: InfoWorld (infoworld.com), American Bazaar Online (americanbazaaronline.com), Meta AI Blog (ai.meta.com).