Skip to content
NewsProduct launch

Meta ships Muse Code, its first coding terminal agent, powered by Muse Spark 1.2

· by Pondero Newsdesk

The short version

Meta released Muse Code in beta on August 5, 2026, a terminal coding agent built on Muse Spark 1.2 that runs persistent background sub-agents throughout each session, a design intended to reduce redundant lookups on long-horizon repository tasks.

Meta ships Muse Code, its first coding terminal agent, powered by Muse Spark 1.2

Meta's first coding terminal agent reached beta on August 5, 2026. Muse Code, built on a new coding-focused model called Muse Spark 1.2, runs persistent background agents throughout an entire session rather than spawning fresh sub-agents per task. Per the Meta AI research blog, that design reduces repeated information-gathering and manual steering on complex, multi-step work.

What

Muse Code enters a market with Claude Code, Codex CLI, and Cursor already established. The agent uses an async background agent system where specialized sub-agents stay active for the full duration of a session, carrying out next steps and reporting back to the main agent only when needed. That is different from architectures that spin up a sub-agent on each task call and tear it down when the call ends.

The runtime logs every model call, tool run, approval, and edit to a local event file, making sessions replay-exact and restart-safe. After a crash, the agent resumes from the exact stopping point. Three skills ship built-in: /plan turns a task into an approval-gated plan, /grill stress-tests that plan until it holds up, and /goal drives toward completing a specified objective. Installation runs on macOS and Linux via a single curl command.

Muse Spark 1.2, the underlying model, updates Muse Spark 1.1 with scaled training compute on coding tasks and expanded training environment diversity. Meta trained it on long-horizon tasks including whole-repository generation and auto-research, and used Muse Spark 1.1 to generate challenging coding environments for a self-improvement loop. Per the announcement, the model was evaluated on Terminal-Bench 2.1, DeepSWE 1.1, and an internal coding benchmark, with a full methodology report published separately.

One case study in the announcement ran the model on GPU kernel optimization over 1,000 or more tool calls across sessions lasting up to 24 hours, demonstrating sustained progress on KDA and MLA kernels for NVIDIA Hopper GPUs.

Muse Spark 1.2 is also available directly through the Meta Model API. Pricing is $1.25 per million input tokens and $4.25 per million output tokens for the standard tier, per the Muse Code product page. A contributor tier, which lets Meta use requests to improve its products, costs $0.10 per million input tokens and $0.20 per million output tokens. Both tiers support a 1 million token context window.

Why it matters

Session-persistent background agents represent a structural departure from how most production coding agents work today. Agents that reacquire context for every sub-task invocation pay a latency and token cost each time. Muse Code's background agents accumulate context once and carry it forward, which matters most on tasks that span many tool calls across a long session.

Pricing positions Muse Spark 1.2 below the Claude Sonnet rate for operators shopping on API cost. With a 1M context window, the model covers large codebases without hitting a hard cutoff. The beta designation is the primary caveat: the API surface may change before general availability, and independent benchmark comparisons are not yet published.

Release cadence adds context. Meta published Muse Spark 1.1 on July 9, 2026, twenty-seven days before this release. That pace is faster than the company's earlier model rollout schedule and consistent with the announcement's statement that "more powerful models" are on the horizon.

What to watch next

Community SWE-bench comparisons against Claude Code, Codex CLI, and Cursor will be the first independent read on where Muse Spark 1.2 actually lands on standard coding evals. Meta's blog showed bar charts for Terminal-Bench 2.1 and DeepSWE 1.1 without publishing numeric scores inline; the full numbers are in the methodology report. Whether the session-persistent background agent pattern spreads to other terminal coding tools in the next 30 to 60 days is the architectural question that follows the benchmark results.

Sources