Meta launches Muse Code, a terminal-based AI coding agent built on Muse Spark 1.2, designed to help developers plan, write, and validate code across large repositories using persistent sub-agents that run concurrently without disrupting existing code.

Pricing is structured with a standard tier at $1.25 per 1 million input tokens and $4.25 per 1 million output tokens, while a Contributor tier offers reduced rates of $0.10 per 1 million input tokens and $0.20 per 1 million output tokens, plus the ability to use prompts and completions for model training.

Bundled commands include /plan to generate an approved plan, /grill to stress-test it, and /goal to drive toward completing the objective; an example interprets an MP4 home video to craft a vacation marketing page.

Investors remain wary of Meta’s high AI-related spending, reflecting ongoing leadership changes within the company’s AI division and Mark Zuckerberg’s 2025 investment moves.

Analysts anticipate Meta’s entry and pricing strategy could challenge Anthropic’s leadership in AI benchmarks by September 2026.

The piece underscores the rapid AI arms race’s strategic importance and risks, with users weighing performance against data protection as AI becomes central across sectors.

Opt-in data training raises enterprise compliance and governance considerations beyond financial benefits, especially for proprietary code and regulated workloads.

Enterprises face practical and legal questions around data training opt-in, making governance a parallel priority to cost savings.

A cautionary note about social-engineering hacks used by OpenAI and Anthropic agents is included as a tip reference but not elaborated in the main article.

U.S. government AI-safety rules are expected to target closed models from OpenAI, Google, and Anthropic, with open Meta models and China’s DeepSeek potentially less affected.

Authorities reportedly plan to regulate closed AI models more stringently, while open models from Meta may face lighter constraints.

AI tools are described as assistive rather than human-like reasoning, with ongoing debates about privacy and trust in Meta software.