Meta AI Chief Alexandr Wang Teases Company's Most Capable Model Yet
Key Takeaways
- •Meta's Chief AI Officer Alexandr Wang announced an imminent release of the company's most capable AI model, positioning Meta for direct competition with OpenAI, Anthropic, and Google.
- •The forthcoming model, codenamed Watermelon, reportedly nears the performance of OpenAI's GPT-5.5 but achieves this while consuming significantly greater computational resources.
- •Meta abandoned its open-weight release approach in favor of proprietary models after Llama 4 drew widespread criticism upon its April 2025 launch.
- •Wang joined Meta in June 2025 after the company made a $14.3 billion investment in his previous firm, Scale AIMeta released Muse Spark version 1.3 in September 2026, delivering improvements in coding performance and overall efficiency.
Meta Platforms' Chief AI Officer Alexandr Wang has announced that the company will soon release its most capable AI model to date. The declaration signals an escalation in Meta's AI ambitions under's leadership and positions the social media giant for a more direct confrontation with OpenAI, Anthropic, and Google in the frontier model race.
Wang framed Meta's recent Muse Spark model, launched in April 2026 and internally codenamed Avocado, as merely an “appetizer” for what is coming next.
From Open-Weight Champion to Proprietary Contender
Meta's strategic shift follows the underwhelming reception of Llama 4 in April 2025, a release that drew widespread criticism and pushed leadership to rethink its entire approach. The move from open-weight releases to a proprietary model strategy marked a philosophical U-turn for a company that had previously positioned itself as the champion of open AI development. In practice, the pivot ends the freely downloadable weights that outside developers and researchers could once build on, aligning Meta's release strategy more closely with the closed approach of the frontier rivals it now faces.
Wang joined Meta in June 2025, following a $14.3 billion investment in his previous company, Scale AI. Under his direction, the company's focus has turned toward building scalable AI systems designed to be woven directly into Meta's apps and hardware ecosystem, a product-first strategy that ties frontier model progress to the platforms Meta already operates.
Watermelon: Meta's Answer to GPT-5.5
The next model in Meta's pipeline carries the codename Watermelon. According to reports, it approaches performance parity with OpenAI's GPT-5.5, though it achieves this by committing significantly more computational resources to the task. That compute-heavy path highlights one of the central tensions in frontier AI development: closing capability gaps can require steep increases in computing power, which places the burden on demonstrating that the added spend translates into measurable performance gains.
Meanwhile, Meta released Muse Spark version 1.3 in September 2026, an update that brought improvements to coding performance and overall efficiency.
Wang has characterized the upcoming releases as “highly competitive” against the best models currently available. When Watermelon arrives, how it measures up against GPT-5.5 and how deeply Meta integrates it into its apps and hardware will offer the clearest read on how far the company's proprietary turn has taken it.
