Meta released Muse Spark 1.2 and Muse Code on August 5, 2026, marking the company's entry into the AI coding-agent market. Muse Spark 1.2 is a coding-focused foundation model purpose-built for debugging sprawling codebases and handling multi-step reasoning tasks, with a 1M-token context window. Muse Code is a terminal-based coding agent that runs on macOS and Linux with integrated planning, subagents, worktrees, and crash recovery via event logging.
Standard API pricing for Muse Spark 1.2 is $1.25 per million input tokens and $4.25 per million output tokens. Meta introduced a new contributor tier at $0.10/$0.20 per million tokens (an 80% discount on input, 95% on output) in exchange for permission to use your prompts and completions for training. The contributor tier has a rate limit of 60 requests per minute versus 3,000 RPM on standard tier.
Muse Spark 1.2 was co-trained with Muse Code to optimize performance within the harness—a deliberate departure from competitors like Anthropic and OpenAI who adapt general-purpose models. Meta reported the model scores second on Terminal-Bench 2.1 after Claude Opus 5, and ranks 49th overall with a 61.5 score. On independent Vals benchmarking under a common harness, Muse delivers the lowest cost-per-test ($0.69) among the top five coding models.
For architects and engineers, Meta's entry into the coding-agent market validates demand at multiple price tiers: premium ($1.25/$4.25) for enterprises, and a training-data-tradeoff tier ($0.10/$0.20) for teams comfortable with data sharing. The co-trained model-plus-harness approach, combined with Muse Glimmer's open-weight 30B release, signals Meta's pivot from open weights in Llama to a hybrid model where closed-weight reasoning models are monetized via API while efficiency-focused open weights remain free.