The 800 mistakes that could reshape Meta’s AI coding strategy

The New Stack · 2 months ago · 39 ● 6 sources

Meta is collecting code fixes from 7,000 engineers using its internal MetaCode AI tool to improve its AI coding models, with over 800 corrections already submitted to refine Muse Spark 1.1 and train a model called Watermelon. The company's Muse Spark 1.1 scored 53% on the DeepSWE leaderboard in July, trailing GPT-5.6 Sol at 73% and Claude Opus 5 at 74%, while Meta's pricing at $1.25 per million input tokens is cheaper than competitors but facing pressure from price cuts and open-weight rivals. By capturing real engineering workflows rather than synthetic benchmarks, Meta aims to close the performance gap with frontier models and reduce costs of running coding agents across thousands of employees.