Meta's new AI, 'Watermelon,' will soon be released with performance comparable to OpenAI's GPT-5.5.

Alexandre Wang, Chief AI Officer at Meta, has revealed that Meta's AI model, Muse Spark, will soon be updated with improvements to its coding and agent capabilities. While the AI model's name itself will remain the same, the internal code will be updated from Avocado to Watermelon.

First, Mark was clearly talking about the industry's progress on agentic capabilities on the whole.

But, while we're on the topic: Our next Muse Spark update is coming soon. Big improvements in coding and agentic capabilities to be more competitive with other leading models.…

https://t.co/uTjx8sZM2A

— Alexandr Wang (@alexandr_wang) July 3, 2026

Meta to release new AI model with advanced coding capabilities 'soon' - SiliconANGLE
https://siliconangle.com/2026/07/03/meta-release-new-ai-model-advanced-coding-capabilities-soon/

Meta AI chief says 'Watermelon' model has caught up to GPT-5.5
https://americanbazaaronline.com/2026/07/03/meta-ai-chief-says-watermelon-model-has-caught-up-to-gpt-5-5-484022/

Meta says its next-gen AI model 'Watermelon' uses 10x more compute, matches GPT-5.5 benchmarks – Firstpost
https://www.firstpost.com/tech/meta-says-its-next-gen-ai-model-watermelon-uses-10x-more-compute-matches-gpt-5-5-benchmarks-14028469.html

In April 2026, Meta announced 'Muse Spark,' the first in its 'Muse' family of AI models aimed at realizing superintelligence for personal use.

Meta announces native multimodal inference model 'Muse Spark' as part of a 'fundamental overhaul' of its AI business - GIGAZINE

According to Mr. Wang, the next model, internally codenamed Watermelon, has achieved scores comparable to OpenAI's GPT-5.5 in multiple benchmark tests. In SWE-Bench Pro , an index for evaluating AI coding agents, the previous version of Muse Spark, internally codenamed Avocado, scored 52.5%, while GPT-5.5 scored 58.6%. GPT-5.5 also scored higher than Muse Spark (Avocado) in the popular coding benchmark Terminal-Bench 2.0, so it appears that Meta has succeeded in closing the performance gap with Watermelon.

Muse Spark has a 'Deliberation Mode' that uses an AI agent to improve the quality of responses to prompts, and Meta's internal tests reported that enabling it improved scores by 8%.

In response to a user's question, 'When will a model comparable to Claude Opus be available?', Wang replied, 'Very soon, very soon.' Anthropic's Claude Opus 4.8 outperformed GPT-5.5 and Muse Spark in some coding tasks, achieving a score of 69.2% in SWE-Bench Pro. However, it did not measure up to GPT-5.5 in Terminal-Bench 2.0.

pretty soon. I think you'll like what we have cooking

— Alexandr Wang (@alexandr_wang) July 3, 2026

Furthermore, there is information suggesting that improvements in Watermelon's output quality are more due to increased infrastructure usage than to improvements in model capabilities.