Meta Superintelligence Labs shipped Muse Code on Wednesday — a terminal coding agent, public beta, macOS and Linux, installed with a single curl line. It runs on Muse Spark 1.2, a coding-focused model Meta says it co-trained with the agent itself. It lands directly opposite Claude Code and Codex.
Vendors do not usually publish the chart where they come second. Meta published three of them.
On Terminal-Bench 2.1, Meta’s own launch deck puts Claude Opus 5 at 86.7% and Muse Spark 1.2 at 82.9%. On DeepSWE 1.1: Opus 5 at 65.0%, GPT-5.6 Terra at 64.8%, Muse Spark 1.2 at 59.3%. On Meta Internal Coding Bench — the benchmark Meta built and controls — Opus 5 posts 79.4% against Muse Spark 1.2’s 70.6%. Meta’s widest loss is on its home fixture.
The version that went around Wednesday night — “59% on DeepSWE, beating Grok 4.5 and Gemini 3.6 Flash” — is true. It is also the narrowest true statement available. Meta shipped a deck it does not win, and shipped it anyway, because the score is not what it is selling.
The price is the product
There is no Muse Code subscription. Billing is per token, on one of two tiers, and the spread between them is the entire launch.
Standard runs $1.25 per million input tokens, $4.25 per million output, $0.15 cached. Meta commits that prompts on this tier are not used to train its models. That is the identical list price it set for Muse Spark 1.1 and its first paid API in July — parity with the field, not below it.
Contributor runs $0.10 input, $0.20 output, $0.002 cached. That is 12.5x cheaper on input, 21x on output, 75x on cached context. The condition, stated plainly by Meta: it may use your prompts and completions to train future models.
Put a realistic agent step through the grid — 60,000 tokens of repository context in, 3,000 tokens of plan-and-patch out. A thousand of those steps costs roughly $87 on the standard tier cold, and about $0.72 on the contributor tier with a warm cache. Meta’s own headline demo, a GPU kernel optimization run of 1,000-plus tool calls over 24 hours on Nvidia Hopper hardware, is ninety dollars of tokens on the tier that protects your code and pocket change on the tier that does not.
Our take: A terminal coding agent’s entire function is to read your codebase. So the prompts on the contributor tier are your source, your internal APIs, your comments explaining why the workaround exists, and whatever your test fixtures happen to hold. That is not a billing preference with a discount attached — it is a data-licensing decision wearing one, and it belongs in front of whoever signs off on data handling rather than whoever watches the cloud bill. Meta set the discount at 12x because the code is worth at least 12x to Meta. The tell is that the discount applies per token, so the incentive to quietly downgrade grows exactly as your usage does.
What Meta actually built
Strip the benchmarks out and the engineering pitch is about not dying. Muse Code appends every model call, tool run, approval and edit to a local event log before execution — “replay-exact and restart-safe,” in Meta’s phrasing. Kill it twenty hours into a job and it resumes where it stopped instead of re-deriving context. Background agents stay alive for the whole session rather than spawning per subtask, and big jobs fan out to sub-agents in isolated git worktrees, so your working copy is never in the blast path.
That log is also an audit artifact: a complete local record of what an autonomous process did to your tree. It is exactly the document a security review asks for and most agent CLIs cannot produce — which lands with some weight this week, given that the previous model in this family reached a third party’s production systems during a safety evaluation.
What to watch
- Independent numbers. Meta published no methodology behind the charts, and nobody outside Meta has run the agent on real repositories yet.
- The event log under someone else’s thousand calls. Reliability is the whole differentiator, and it is currently unverified beta software.
- Where contributor-tier keys turn up. Watch whether they start appearing inside proprietary repos, quietly, on the strength of the price.
- The platform gap. macOS and Linux only, terminal only, no IDE extension, no Windows build. For enterprise fleets that is a constraint, not an inconvenience.
If you run long-horizon jobs on a scratch clone, Muse Code is worth an afternoon. If you run a team against a proprietary repository, nothing here forces a move — and before you hand any agent write access to anything, run the blast radius checklist first.