Meta is the fourth major lab in the terminal coding agent race. On the evening of August 5, 2026, Mark Zuckerberg announced Muse Code in beta — a terminal agent built for what AI at Meta calls "long-horizon software engineering," powered by the new Muse Spark 1.2 model. Alexandr Wang, who heads Meta Superintelligence Labs, confirmed it's the first coding agent from that org: "muse code in beta is live. first coding agent from msl, built on muse spark 1.2."
That puts Muse Code in the same bucket as Claude Code, OpenAI's Codex, and Google's Antigravity CLI. Install is one line:
curl -fsS https://dev.meta.ai/install.sh | bash
The interesting part isn't that Meta shipped a CLI. It's the architecture, and the pricing model underneath it — one of which I'd adopt, and one of which deserves a hard look before you point it at proprietary code.
The Worktree Fan-Out and the Event Log
Muse Code's two distinctive mechanisms are both answers to real problems with long-running agents.
First, it keeps specialized background agents alive for your whole session, so context — files inspected, decisions made, repo patterns learned — carries forward instead of resetting on every task. When a job is big enough, it fans out to sub-agents working in parallel, each in its own isolated Git worktree. "Your working copy is never touched," per Zuckerberg. He demoed it building six features for a game simultaneously with no collisions. OpenCode's desktop client uses the same worktree mechanism, so this isn't novel, but it's the right primitive for parallel agent work: multiple branches checked out to separate directories, no merge fights in your main branch.
Second, the crash recovery pattern. Every model call, tool run, and edit hits a local event log before it executes. If the agent crashes mid-task, it replays from that log and resumes exactly where it left off — no lost work, no re-prompting you to explain the task again. This is a write-ahead log applied to agent actions, and it's a direct fix for the most common complaint about goal-mode agents: a dropped session used to mean starting over. I'd genuinely like to see this become table stakes across harnesses.
The stress test was a 24-hour unattended run — 1,000+ tool calls on NVIDIA Hopper pointed at a kernel optimization task, still finding improvements past the initial exploration phase. Vendor-reported, like the six-feature demo, so treat it as a claim until third-party reproductions exist.
Where It Actually Sits on Benchmarks
Meta published its own comparison chart, and the headline is honest: Muse Spark 1.2 trails Claude Opus 5 on every single benchmark Meta chose to publish.
| Model (harness) | Terminal-Bench 2.1 | DeepSWE 1.1 | Meta Internal |
|---|---|---|---|
| Opus 5 (Claude Code) | 86.7% | 65.0% | 79.4% |
| Muse Spark 1.2 (Muse Code) | 82.9% | 59.3% | 70.6% |
| GPT 5.6 Terra (Codex) | 81.8% | 64.8% | 65.4% |
| Grok 4.5 (Grok Build) | 81.6% | 56.6% | — |
| Gemini 3.6 Flash (Antigravity CLI) | 78.9% | 40.0% | 63.9% |
So: behind the frontier leader, but it beats GPT 5.6 Terra and Grok 4.5 on Terminal-Bench 2.1, and it's a clear jump over its own predecessor Muse Spark 1.1 (76.2% → 82.9% on Terminal-Bench 2.1). These are Meta's own self-selected comparisons at "max"/"high" settings, so validate on your own repo before switching tooling. A four-point gap to Opus 5 on Terminal-Bench isn't noise — for hard multi-file work I'd still default to Claude Code today.
The Contributor Tier Is Priced in Data
Here's the part that changes the calculus. Zuckerberg's pricing line is "easy and low-cost to get started... on our contributor tier." Per public commentary around the launch — not a confirmed Meta pricing page — the contributor tier runs roughly $0.10 per million input tokens versus about $1.25/M on the standard tier. The trade: Meta gets to use your usage data for model training.
That's a ~92% discount priced in data, not dollars — the familiar free-tier bargain, made explicit. Before you run Muse Code against a proprietary codebase, confirm the exact terms and toggle settings at dev.meta.ai. And note the open-source question is unresolved: asked directly on X whether Muse Code would be open source, Zuckerberg's reply was "I'll have more to share on that soon." Muse Spark 1.2 is closed regardless of what happens to the harness.
What I'd Watch Before Pointing This at Production
Concrete reasons to hold off: it's beta, the two headline demos are unverified, the pricing specifics are secondhand, and the data-training trade on the cheap tier is a real governance question for anything sensitive. But the architecture — persistent agents, worktree isolation, write-ahead event logging — is directionally where the category should head. I'd keep Muse Code on the shortlist to re-evaluate when it leaves beta and when independent numbers land. And if the $0.10/M tier holds up for non-proprietary work, it starts to look genuinely interesting against a $1.25/M standard tier elsewhere.
Sources: