Meta Launches AI Coding Agent Muse Code With Controversial Data-Training Pricing Tier, Trails Claude Opus 5 on Benchmarks

Meta launches Muse Code, a beta AI coding agent with parallel background agents, powered by Muse Spark 1.2 which scores second behind Claude Opus 5 on benchmarks, but sparks controversy with a two-tier pricing model that lets users pay less in exchange for allowing Meta to train on their private code.

• Meta launches Muse Code, a terminal-based AI coding agent now in beta, featuring persistent async background agents and parallel worktrees that allow multiple features to be built simultaneously without touching the developer's working copy.
• Muse Spark 1.2, a coding-focused model update co-trained with Muse Code itself, scores 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1, placing it in strong second behind Anthropic's Claude Opus 5, which leads all three benchmark charts.
• Meta introduces a controversial two-tier pricing model where a cheap 'contributor tier' at $0.10 per million input tokens grants Meta permission to train on users' code and prompts, while a standard tier at $1.25 per million input tokens keeps data private — with no open-source model weights released despite Meta's prior reputation as an open-source AI champion.