Meta Ships Muse Code: An Event Log, Git Worktrees, and a Data-for-Discount Tier

On the evening of August 5, 2026, Mark Zuckerberg posted that Meta had a coding agent. The next morning the detail was public: Muse Code, in beta, running on a new coding-focused model called Muse Spark 1.2. It installs with one command, lives in the terminal, and is aimed squarely at the ground Anthropic's Claude Code and OpenAI's Codex have occupied for the past year.

Two things here are worth a working engineer's attention, and neither is the benchmark chart. One is an architectural decision about how the agent records what it did. The other is a pricing decision about who owns the code it sees. The rest is a competent entry in a crowded category.

What Meta shipped

Meta's launch post describes Muse Code as taking on “complex software engineering tasks across large repositories: planning changes, writing code, and validating the results.” It runs on macOS and Linux, installs via curl -fsSL https://dev.meta.ai/install.sh | bash, and reaches the model through the Meta Model API. SiliconANGLE reports it is also available through OpenRouter.

The model underneath, Muse Spark 1.2, is a coding-focused update to the Muse Spark 1.1 release we covered when Meta started charging for its models in July. Meta claims improvements in “code generation, complex debugging, codebase understanding, and end-to-end developer workflows,” and says the model was trained on long-horizon work including whole-repository generation, with rejection sampling and self-improvement loops for instruction-following. It keeps the 1M-token context window of its predecessor, which is the size class we walked through in our million-token context explainer.

The framing throughout is scale. This is not pitched as a tool for writing a function; it is pitched at repositories. In Meta's own telling, one internal test had the agent making more than 1,000 tool calls over as long as 24 hours on a GPU kernel task.

How the agent is built

Three design choices carry the release.

The event log. Meta says every model call, tool run, approval and edit is appended to a local log, making the session “replay-exact and restart-safe.” If the process dies four hours into a long task, it resumes from the log rather than restarting. That is the genuinely useful idea here, and its second-order value is bigger than crash recovery: an agent that writes a complete, ordered record of its own actions is an agent you can audit after the fact. Most coding agents today leave you a scrollback buffer and a diff. A durable log is the difference between “the agent changed these files” and “here is every decision, tool call and approval that produced this change, in order.”

Persistent background agents. Rather than spawning a helper per task and tearing it down, Muse Code keeps a set of async background agents alive for the whole session. They accumulate context as they go and report back to the main agent. The claimed benefit is less latency and less repeated information-gathering on multi-step work.

Worktree isolation. When a job is large enough to split, Muse Code fans out to sub-agents that each check out into a separate git worktree. Your working copy is never touched. Zuckerberg's demonstration example, as reported by TechCrunch, was building six features for a game simultaneously with no collisions.

Muse Code also ships three default skills: /plan produces an approval-gated plan, /grill stress-tests that plan, and /goal works toward completion of an objective.

Which parts are actually new

Feature	What it does	How novel
Append-only event log	Records every call, tool run, approval and edit; enables exact replay and restart	The strongest idea in the release; auditability as a first-class property
Persistent background agents	Helpers stay alive across the session and accumulate context	An engineering refinement of the usual spawn-per-task pattern
Worktree isolation	Parallel sub-agents each get their own checkout	Built on a plain git primitive you can already script yourself
Bundled plan / grill / goal skills	Approval-gated planning and plan stress-testing	Comparable patterns exist across current agents

Worth saying plainly: git worktree has been in git for a decade. Any agent you already run can be pointed at a detached worktree, and several teams do exactly that. If parallel isolation is what you want, you do not need to switch tools to get it. Our field guide to coding agents and the four-way tool comparison cover how the current crop differ on the parts that are harder to replicate.

The price, and what the cheap tier costs

Meta's launch post does not publish rates. The figures below come from launch coverage by Engadget and MacRumors, which agree.

Reported Muse Spark 1.2 tiers

Tier	Input / output per 1M tokens	Reported rate limits	Data terms
Standard	$1.25 / $4.25	3,000 requests/min, 4M tokens/min	No training use
Contributor	$0.10 / $0.20	60 requests/min, 2.1M tokens/min	Meta may use your prompts and completions to train future models

That is 12.5x cheaper on input and 21x cheaper on output. Meta chief AI officer Alexandr Wang, quoted by Engadget, described it as “a contributor tier that gets you in at a significantly lower cost,” and told CNBC, per TechCrunch's account, that “for a lot of workflows and a lot of use cases, this can be an incredibly good option, especially from a cost perspective.”

Here is the part to sit with. The contributor tier is not a pricing decision a developer makes; it is a data-governance decision an organization makes. A coding agent pointed at a repository reads far more than the file you asked about. It reads configuration, internal service names, schema, comments, credentials that should not be in the repo but sometimes are, and the shape of your architecture. Sending all of that into a training corpus is a different act than sending a prompt to an endpoint that discards it.

For a hobby project, the discount is close to free money. For code covered by a client NDA, for anything touching controlled unclassified information, for a codebase whose value is partly that competitors cannot see it, the calculus inverts. If your work sits in that second category, the relevant reading is our primer on CUI for software teams and the broader piece on what changes when AI touches regulated data. Federal buyers have already started writing this question into contract language, which we covered when GSA published its LLM data-safeguarding clause.

The practical instruction is short: before the agent touches a real repository, verify which tier your account is on, and write the answer down somewhere your security reviewer can find it.

Reading the benchmark chart

Meta's post charts three evaluations: Terminal-Bench 2.1, DeepSWE 1.1 (113 tasks across 91 repositories in five languages), and an internal coding benchmark of 440 tasks drawn from real pull requests. The Terminal-Bench figure that circulated in launch coverage put Muse Spark 1.2 in Muse Code at 82.9%, ahead of GPT-5.6 Terra in Codex at 81.8% and Grok 4.5 in Grok Build at 81.6%, and behind Claude Opus 5 at max effort in Claude Code at 86.7%.

Two cautions before anyone quotes that in a planning meeting.

First, read the labels. Every one of those entries names a model and a scaffold. “Opus 5 at max effort in Claude Code” is not a model score; it is a model-plus-scaffold-plus-effort-setting score. Swap the scaffold and the number moves. That is a fact about agents rather than a flaw in the benchmark, and it means a leaderboard row cannot tell you how a model will behave inside your setup. It is why teams need their own evaluation rig, the subject of our guide to how teams test agents in 2026.

Second, vendor-run and independently-run numbers differ. The independent evaluator Artificial Analysis published its own measurements the same day: Terminal-Bench v2.1 at 80%, an Intelligence Index of 54, and a position it describes as effectively tied with GPT-5.5 and Grok 4.5, narrowly behind Claude Opus 5, Claude Fable 5, GPT-5.6 Sol and Kimi K3. Different tests, different conditions, so the gap is not evidence of anything improper. It is evidence that a launch-day chart is a marketing artifact and a third-party measurement is a data point.

What to do with this

The following is our analysis, not reported fact. Most teams should not switch tools this week; a beta agent from a new entrant is not where production work belongs. Four things are worth doing anyway.

Take the event-log idea, whether or not you take the tool. If your current agent does not leave a durable, ordered record of its actions, that is a gap you can close yourself by capturing tool calls and diffs into an append-only file. When someone asks in three months why a function changed, a log answers and a scrollback does not. This is the same instinct behind LLM observability: instrument first, investigate later.

Re-price your token budget. Meta has now twice used price as its point of entry, and a well-capitalized company that does not need API margin sets a floor everyone else has to answer. Whether or not Muse Spark ships in your stack, its rate card is bargaining power in your next renewal. Our coding-tool pricing guide and the context-economics piece cover where the money actually goes on agentic workloads: usually output tokens and cache misses rather than the headline input rate.

Decide the data question once, at the organization level. The contributor tier will not be the last time a vendor offers a discount for training rights, and the worst outcome is each engineer answering it individually inside their own terminal. Write a one-paragraph policy: which repositories may reach a training-eligible endpoint, and which may not.

Watch the open-weights question. Asked about open-sourcing Muse Spark, Zuckerberg said he would have more to share soon, per The Register. Meta built its AI reputation on giving weights away and has spent 2026 selling access instead. If that reverses for a model of this class, the calculation for teams that cannot send code off-premises changes substantially, in a market where open-weight coding models are already competitive.

Sources: Meta — “Introducing Muse Code and Muse Spark 1.2” (August 5, 2026); TechCrunch; Engadget; MacRumors; The Register (August 6, 2026); SiliconANGLE; Artificial Analysis — Muse Spark 1.2; MarkTechPost. Pricing figures are secondhand reporting, not lifted from a Meta price page; confirm current rates in your own console. Analysis and framing by Precision AI Academy.

Common questions

Can I use Muse Code on Windows? Not directly at launch. Coverage of the beta lists macOS and Linux.

Does the contributor tier cost me anything besides data? Throughput. The reported rate limits are 60 requests per minute and 2.1M tokens per minute, against 3,000 requests and 4M tokens on standard. For a long agentic run that difference is real.

Is Muse Spark 1.2 better than Claude Opus 5 or GPT-5.6? On the launch-day Terminal-Bench comparison it landed second of four, behind Opus 5. Independent measurement from Artificial Analysis places it a step further back. The honest answer is that it is in the same class, and the ranking depends on the test and the scaffold.

Should I move production work onto it? It is a beta. Run it on something you can afford to throw away first, and measure it against your own tasks rather than a published chart.