Meta has launched Muse Code, its first coding agent, alongside Muse Spark 1.2, the model that powers it. Announced by Mark Zuckerberg on 5 August, the terminal tool installs with one command and handles complete engineering tasks — planning changes, writing code and validating results — across entire repositories.
Muse Code’s pitch is its multi-agent architecture. Specialised background agents stay active for the whole session, building context instead of starting from scratch, and big jobs fan out to parallel subagents working in isolated git worktrees — Zuckerberg says testing had it building six features for a game at once with no collisions. Every model call, tool run and edit is written to a local event log before it executes. His framing for that:
If it crashes mid-task, it picks up exactly where it left off from that log. No lost work and no re-prompting.
The bundled skills are named like a bootcamp: /plan turns a task into an approval-gated plan, /grill stress-tests it, /goal drives it to completion.
Muse Spark 1.2 is a coding-focused update to last month’s 1.1, with a 1M-token context window and improvements in code generation, debugging and codebase understanding. It was co-trained with Muse Code itself, and Meta says it generalises to other coding agents. Meta’s research post shows it competitive on Terminal-Bench 2.1 and DeepSWE 1.1, and one case study had it optimising NVIDIA GPU kernels over more than 1,000 tool calls across 24 hours.
Pricing: the contributor tier, and what it costs you
Pricing is where Muse Code differs from the rest of the field. The contributor tier — muse-spark-1.2-contributor — costs about AED 0.37 per 1M input tokens and AED 0.73 per 1M output (USD 0.10 and USD 0.20), rate-limited by tokens in a rolling five-hour window rather than by request count. The trade is stated plainly: prompts and completions may be used to train Meta’s models.
The standard tier is pay-as-you-go at about AED 4.59 per 1M input and AED 15.61 per 1M output (USD 1.25 and USD 4.25), with cached input at about AED 0.55 per 1M (USD 0.15). Prompts are not used for training on this tier, and zero data retention is available on request.
Availability
Muse Spark 1.2 is available today in Muse Code, the Meta Model API and OpenRouter, with expanded global access. The contributor tier is limited to select countries, so UAE developers should check whether their region qualifies; the standard tier is open. The launch pits Meta directly against Anthropic’s Claude Code and OpenAI’s Codex — a market that turned on Anthropic’s pricing earlier this year, when Claude Code moved to $100+ plans and developers pushed back.
It is the coding-focused follow-up to the frontier model Meta rolled out across its apps earlier this year. Muse Code is a beta, and Meta is open that larger models are coming.
What makes Muse Code different from other coding agents?
It runs persistent background agents that build context over a whole session, fans large jobs out to parallel subagents in isolated git worktrees, and writes every model call, tool run and edit to a local event log — so a crash mid-task resumes from the log instead of restarting.
How much does Muse Code cost?
Two tiers. The contributor tier is about AED 0.37 per 1M input tokens and AED 0.73 per 1M output (USD 0.10 and USD 0.20), rate-limited by tokens in a rolling five-hour window, in exchange for allowing Meta to use your prompts to train its models. The standard tier is pay-as-you-go at about AED 4.59 per 1M input and AED 15.61 per 1M output (USD 1.25 and USD 4.25), with cached input at about AED 0.55 per 1M (USD 0.15); prompts are not used for training.
Is Muse Code available in the UAE?
Yes. Muse Spark 1.2 is available globally in Muse Code, the Meta Model API and OpenRouter, and Muse Code installs with a single command on macOS and Linux. The cheaper contributor tier is limited to select countries, so UAE developers should check whether their region qualifies.
Leave a Reply