Meta Muse Code: Terminal Coding Agent on Muse Spark 1.2

Introduction

Meta Superintelligence Labs released Muse Code, a terminal-based coding agent, in beta on August 5, 2026, alongside Muse Spark 1.2, an updated version of the company's proprietary large language model tuned specifically for software engineering. On Meta's own evaluation harness, Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1, up from 76.2% for the prior version, placing second behind Claude Opus 5 among the six frontier models Meta compared. Muse Code is the delivery vehicle for that model: it plans, writes, and validates code changes across large repositories directly from the command line, with no graphical app required.

The launch marks Meta's third Muse Spark iteration in about four months, following the original Muse Spark model in April and Muse Image in July, and signals an aggressive push to compete with Anthropic's Claude Code, OpenAI's Codex, and Google's coding agents in the fast-growing terminal-agent category.

Feature Overview

Muse Code introduces several architectural features aimed at long-running, multi-step coding sessions rather than single-shot code generation:

• Async background agents: specialized agents persist for the duration of a session instead of being spun up per task, cutting latency and avoiding redundant context-gathering on follow-up requests.
• Crash-safe event log: a local log records every model call, tool run, approval, and edit as it happens, so a crash or interruption can resume from the last verified state instead of losing work.
• Skill commands: built-in /plan (produces an approval-gated plan before execution), /grill (stress-tests a plan for weaknesses before committing to it), and /goal (drives the agent toward a stated end-state) give users explicit control points in an otherwise autonomous workflow.
• Worktree isolation: parallel sub-agents each operate in their own git worktree, so simultaneous edits to the same repository don't collide with one another or with the user's working copy.

On the model side, Muse Spark 1.2 posted gains across all three benchmarks Meta reports: Terminal-Bench 2.1 rose from 76.2% to 82.9%, DeepSWE v1.1 from 53.0% to 59.3%, and Meta's Internal Coding Bench from 68.3% to 70.6%. Meta says the model was co-trained with Muse Code and given extended exposure to long-horizon, whole-repository tasks, including a kernel-optimization case study that showed sustained gains over baseline on NVIDIA Hopper KDA and MLA kernels across more than 1,000 tool calls.

Usability Analysis

Installation is a single command (curl -fsSL https://dev.meta.ai/install.sh | bash), and the tool currently supports macOS and Linux only — there is no Windows build at launch. Unlike Claude Code and ChatGPT Codex, which both ship companion app or IDE integrations, Muse Code has no dedicated graphical interface; it lives entirely in the terminal.

Access runs through the Meta Model API on two pricing tiers: a standard rate of $1.25 per million input tokens and $4.25 per million output tokens, and a contributor tier — a 12.5x discount at $0.10 per million input tokens and $0.20 per million output tokens — for users who opt into Meta's data-sharing program. Model weights remain closed, so there is no self-hosting option.

Meta itself flags a caveat worth taking seriously: when comparing Muse Spark 1.2 against Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, Grok 4.5, and Kimi K3, the company notes its evaluation setup, including agent tools and system prompts, "may not be tuned to the strengths of third-party models." More significantly, every published number is vendor-run — no independent Terminal-Bench leaderboard entry existed for Muse Spark 1.2 at launch, and Meta's previous model scored 3.8 points below its own published figure once outside evaluators reproduced the test.

Pros & Cons

Pros

• Meaningful benchmark gains across all three of Meta's coding evaluations compared to Muse Spark 1.1
• Crash-safe event log and worktree isolation address real reliability problems in long-running, multi-agent coding sessions
• Contributor-tier pricing is substantially cheaper than the standard rate for eligible users
• Terminal-first design slots into existing developer workflows without forcing a new app

Cons

• All benchmark numbers are self-reported on Meta's own harness with no independent verification, and the predecessor model missed its published score by 3.8 points once checked externally
• No Windows support and no dedicated app or IDE integration, unlike Claude Code and ChatGPT Codex
• Closed model weights mean there is no self-hosting option
• Still in beta, so stability and feature completeness remain unproven at scale

Outlook

Shipping a third Muse Spark iteration in roughly four months shows Meta treating coding-agent cadence as a competitive necessity, not a side project. Whether Muse Code earns real developer adoption will likely hinge less on the internal benchmark race and more on two concrete gaps: independent verification of the Terminal-Bench and DeepSWE scores, and closing the usability distance to rivals that already ship Windows support and a dedicated app. Third-party benchmark reproductions in the coming weeks should show whether the 82.9% figure holds up outside Meta's own test harness.

Conclusion

Muse Code is a credible, if still rough, entry into the terminal coding agent category, backed by genuine architectural work — crash-safe logging, worktree isolation, approval-gated planning — and a real benchmark jump for Muse Spark 1.2. But the numbers come from Meta's own harness, and the beta ships without Windows support or a GUI companion. It suits macOS/Linux developers willing to test an early but fast-iterating agent; teams wanting a polished, cross-platform daily driver today should wait for independent verification and broader platform support.

Editor's Verdict

Meta Muse Code: Terminal Coding Agent on Muse Spark 1.2 brings real, demonstrable value, though with caveats that deserve weighing.

The strongest case for paying attention is the crash-safe event log and worktree isolation address real reliability problems in long-running, multi-agent coding sessions, which raises the bar for what readers should now expect from peers in this space. Reinforcing that, meaningful benchmark gains across all three of Meta's coding evaluations compared to Muse Spark 1.1 adds practical value rather than just headline appeal. On the other side of the ledger, all benchmark numbers are self-reported on Meta's own harness with no independent verification is a real constraint, not a marketing footnote, and it should factor into any serious decision. Layered on top of that, no Windows support and no dedicated app or IDE integration narrows the set of teams for whom this is an obvious yes.

For product teams, content creators, and knowledge workers looking to upgrade a specific workflow, a measured trial makes sense, with clear criteria for when to expand or pull back. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.