Meta released Muse Code on August 5, 2026, a beta terminal-based coding agent built around its newly announced Muse Spark 1.2 model, according to the company’s AI Research blog. The launch pushes Meta deeper into a crowded field of AI coding agents that already includes Anthropic’s Claude Code, OpenAI’s Codex and Google’s Gemini CLI.
What Muse Code does
Muse Code runs from the terminal and installs on macOS and Linux with a single command. Meta says it uses persistent “async background agents” that stay alive for an entire session rather than spawning fresh for each task, and it keeps a local, append-only event log of every model call, tool run, approval and edit so a session can be replayed or restarted after a crash. The agent ships with three built-in skills: /plan, which drafts an approval-gated plan before making changes; /grill, which stress-tests that plan; and /goal, which works a task through to completion. In one case study, Meta said the agent made more than 1,000 tool calls over as long as 24 hours while optimizing GPU kernels for Nvidia Hopper hardware.
Benchmarks, but light on methodology
Meta said Muse Spark 1.2 — co-trained alongside Muse Code — improves on code generation, debugging and whole-codebase understanding, and it published results on Terminal-Bench 2.1, DeepSWE v1.1 and an internal 440-task benchmark drawn from real pull requests, comparing itself against Claude Opus 5, GPT-5.6, Gemini 3.6 Flash, Kimi K3 and Grok 4.5. However, Meta released those results as chart images without a detailed methodology write-up, Forbes reported, so the comparisons cannot yet be independently verified.
Pricing
Access to Muse Spark 1.2 through Muse Code runs on a “Standard” tier priced at $1.25 per million input tokens and $4.25 per million output tokens, matching what Meta charged for Muse Spark 1.1, per Forbes’ review of Meta’s pricing page. A discounted “Contributor” tier — $0.10 per million input tokens and $0.20 per million output — is available to developers who agree to share usage data with Meta.
Muse Spark 1.2 is the third release in four months from Meta Superintelligence Labs, the unit Meta formed to pursue what it calls personalized “superintelligence,” and follows Muse Spark 1.1’s debut as a standalone coding model in July. It also arrives days after Y Combinator open-sourced QM, its own internal AI agent harness — another sign that vibe coding tools are multiplying faster than developers can evaluate them.