By ByteBot
Meta dropped a terminal coding agent on August 5, 2026. Muse Code is built on Muse Spark 1.2, installs in one command, and competes directly with Claude Code and Codex. The honest verdict: it ranks third on every major benchmark. But it has one architectural feature that neither rival offers, and a pricing tier aggressive enough to make you do the math.
What Muse Code Actually Is
Muse Code is a terminal-based AI coding agent built for complex tasks across large repositories. It plans changes, writes code, and validates results. It runs on macOS and Linux and installs with one command:
curl -fsSL https://dev.meta.ai/install.sh | bashIt ships with three built-in workflow commands: /plan creates an approval-gated task plan before writing a single line of code, /grill stress-tests that plan for edge cases, and /goal drives execution toward a defined objective. The model underneath — Muse Spark 1.2 — was co-trained with the harness itself, which matters for how well the agent and model coordinate on multi-step tasks.
The Benchmark Reality
Meta’s self-reported numbers put Muse Spark 1.2 at 82.9% on Terminal-Bench 2.1. Independent verification tells a more complicated story. Meta’s previous 1.1 model had a 3.8-point gap between its self-reported score (80.0%) and the publicly verified result (76.2%). The jump from 1.1 to 1.2 also includes harness improvements, not just model gains. Until Muse Spark 1.2 lands on the public leaderboard, treat 82.9% as a ceiling estimate.
| Benchmark | Muse Spark 1.2 | Claude Opus 5 | GPT-5.6 Terra |
|---|---|---|---|
| Terminal-Bench 2.1 | 82.9% | 86.7% | 81.8% |
| DeepSWE v1.1 | 59.3% | 65.0% | 64.8% |
| Meta Internal Bench | 70.6% | 79.4% | N/A |
On all three benchmarks, Muse ranks third. If benchmark performance is your primary criteria, Claude Code and Codex both win right now. But benchmarks measure specific task profiles at a point in time. They do not measure what happens when your agent crashes at hour 22 of a 24-hour run.
The Feature Worth Paying Attention To
Muse Code keeps a local append-only event log. Every model call, tool run, approval, and edit gets written to that log before it executes. If the agent crashes mid-task — network failure, machine restart, anything — it resumes exactly where it stopped. Meta calls this “replay-exact and restart-safe.”
The proof case from the launch: 1,000+ tool calls over 24 hours on NVIDIA Hopper GPU kernel optimization, fully resumable after any interruption. The same session ran six parallel game feature builds using isolated git worktrees, with no conflicts. TechCrunch’s coverage notes this crash-safe runtime is the technical edge that distinguishes Muse Code from rivals.
The architecture behind this: Muse Code runs persistent async background agents that stay alive for the full session and accumulate context instead of re-gathering it per task. Rivals spawn fresh context on each handoff. For a 10-minute task, that gap is irrelevant. For a 10-hour refactor, it is not.
There is also an enterprise angle Meta has not pushed loudly enough: the append-only log is a compliance audit trail. A verifiable record of every action the agent took is a requirement in regulated workflows, not a nice-to-have.
Pricing: Two Very Different Tiers
Standard pay-as-you-go: $1.25 input and $4.25 output per million tokens — competitive with the field. Then there is the contributor tier: $0.10 input and $0.20 output, a 12.5x and 21.25x reduction, in exchange for Meta training future models on your prompts and completions.
Cost per solved task tells the full story. VentureBeat’s analysis puts Muse Spark at $0.58 per task, versus $0.82 for Claude Code on Opus 4.8 and $5.57 for Codex. At standard rates, Muse Code is the cheapest option per completed task. At contributor rates, it is effectively in free-tier territory.
The question to ask before opting into contributor pricing: what code are you writing, and who owns the IP? If the answer involves client work, proprietary systems, or anything that should not appear in a training dataset, contributor pricing is not for you regardless of the discount.
Who Should Switch
Try Muse Code if you run long-horizon agentic tasks where crash recovery would save you real hours. Try it if you are cost-constrained and comfortable with the data-sharing trade-off on contributor pricing. Try it if you need a built-in audit trail for compliance reasons.
Stick with Claude Code or Codex if benchmark performance on agentic coding tasks is your deciding factor. Muse Spark 1.2’s verified leaderboard position is an open question. The 1.1 track record suggests some skepticism is warranted until that number appears independently.
Muse Code is a first-version terminal agent from a company that is not primarily a developer tools company. That is not a reason to dismiss it. First versions with genuinely novel architecture tend to matter — the crash-safe replay log and the persistent sub-agent model are ideas worth watching, regardless of where this month’s benchmarks land.
