Meta shipped Muse Code on August 5 with two price lists for the same model. The standard tier bills $1.25 per million input tokens, $0.15 for cached input, and $4.25 per million output, and Meta says prompts and completions sent at that tier are not used to train its models. The contributor tier bills $0.10, $0.002, and $0.20 for the same three things, in exchange for permission to train on what you send. Output tokens cost roughly 21 times less on the cheap tier, cached input about 75 times less.
The agent wrapped around those prices breaks no new ground. Muse Code is a terminal binary for macOS and Linux, installed with a single curl command, running a coding model called Muse Spark 1.2 that Meta says it co-trained with the harness itself. It plans a change, edits files, runs shell commands, checks its own work, and fans large jobs out to background sub-agents in isolated git worktrees. “Your working copy is never touched,” Mark Zuckerberg wrote, describing a test in which six game features were built simultaneously without collisions. Every model call, tool run, approval, and edit lands in a local append-only event log, so an interrupted task can resume without redoing finished steps. There is no desktop app and no IDE plugin. Claude Code and Codex both got here first, and Meta is not pretending otherwise.
The pricing page is where this launch differs from the others. A frontier lab has now published a per-token price for training rights on private source code, and it is low enough that the trade will look obvious to a lot of people.
The two tiers, side by side
| Standard | Contributor | |
|---|---|---|
| Input / Mtok | $1.25 | $0.10 |
| Cached input / Mtok | $0.15 | $0.002 |
| Output / Mtok | $4.25 | $0.20 |
| Rate limit | 3,000 req/min, 4M tok/min | 60 req/min, 2.1M tok/min |
| Meta trains on your data | No | Yes |
The rate limits deserve as much attention as the prices. Sixty requests a minute is one per second across everything the agent is doing, and the headline feature is fanning a task out to parallel sub-agents that each make their own tool calls. Meta’s own numbers say which tier it built for production work: standard gets fifty times the request budget and roughly double the token throughput. The cheap tier is priced for individuals and small experiments, then rate-limited to keep it that way.
Work out what you are actually saving
A 21x ratio is a persuasive number, and it is doing most of the work in the coverage. Run a month of real usage through it and the totals come out smaller than the ratio suggests.
Take one developer running the agent hard. Assume 200 million input tokens a month at an 80 percent cache hit rate, plus 20 million output tokens — the usage figures are an assumption, the prices are Meta’s. On standard that comes to about $159. On contributor it comes to about $8. The saving is roughly $150 a month, or a little under $1,900 a year, for one seat. A five-person shop saves under ten thousand dollars a year, and that assumes all five can share a 60-requests-per-minute ceiling without tripping over each other.
What you hand over for that saving is a permanent training license over whatever the agent decides to read.
Meta deserves credit for how it has done this. The discount is real, the exchange is disclosed up front, and the alternative that costs more is documented rather than buried. Set against the industry habit of acquiring training rights through a terms-of-service revision nobody reads, posting a number is a step forward. The difficulty is that the number is attached to a category Meta has not defined.
”Prompts and completions” means something else in a terminal
In a chat product your prompt is what you typed. In an agentic coding loop it is whatever the agent decided to read: file contents, diffs, test output, stack traces, dependency manifests, config, commit messages, and whatever each sub-agent pulled into its own context on the way to an answer. You do not assemble that payload; the agent does, at machine speed, with a million-token context window to fill.
Meta has not published where the boundary sits. Alexandr Wang, who runs Meta Superintelligence Labs, said the tier is opt-in and declined to give specifics on scope or retention. Nothing in the launch coverage answers whether a file the agent read into context counts as part of the prompt. No retention period for contributed sessions has been stated either. Meta says it is now taking requests for a zero-data-retention option, but has not said how that differs in substance from the standard tier’s no-training commitment. Until those are written down, “prompts and completions” is a term of art carrying more weight than it can hold.
There was a version of this argument in July, when Grok Build turned out to be uploading whole Git repositories, history and committed secrets included, regardless of what the agent had actually read. The lesson was that training consent, transmission consent, and retention are three separate promises that vendors habitually collapse into one switch. Muse Code inverts that shape: here the transmission is the agreed part and the scope is the undefined one. That raises the stakes on getting an answer, because Meta can delete a dataset but it cannot unlearn a model that has already trained on one.
Who can actually grant this license
If you own everything the agent will read, take it. Open source work, permissively licensed code, personal projects, synthetic test repos, throwaway spikes, learning exercises: contribute all of it and run on some of the cheapest coding tokens on the market. For that case the tier is a straightforwardly good deal, and a public price on training data is useful information for everyone.
If someone else owns it, decline. Client code under NDA, unreleased product logic, work covered by a confidentiality clause, repositories touching regulated data, and any dependency whose license you have not read against model training all fall on that side. The question is not whether you trust Meta with the bytes; it is whether you have standing to grant the license at all, and on most agency and contract work you do not.
That last part is where small shops get caught. Most master services agreements written before 2023 say nothing about model training, and the absence of a prohibition is not a permission. If your contract says client materials are confidential and will be used only to deliver the work, routing them into a training pipeline to save $150 a month is not a call you get to make on the client’s behalf. Get it in writing, or stay on standard and treat the difference as insurance.
The operational side is harder than it sounds, because you cannot rely on being careful in the moment. The agent picks what to read. If contributor billing is attached to a key you also use for client work, one stray cd into the wrong directory is all it takes to put NDA material into a training set. Separate keys, separate machines, and a written policy beat good intentions. The local event log helps here, since it gives you an after-the-fact record of what the agent touched, which is more auditability than most agent CLIs offer.
What the benchmark chart actually shows
Meta’s own chart does not put Muse Spark 1.2 in first place. On Terminal-Bench 2.1 it puts Claude Opus 5 at 86.7 running inside Claude Code, Muse Spark 1.2 at 82.9 inside Muse Code, and GPT-5.6 Terra at 81.8 inside Codex. Each model ran in its vendor’s own agent, which measures the pairing rather than the model, and Meta conceded its setup may not be tuned for third-party systems. Treat the whole chart as vendor-reported. Neither Muse Spark 1.2 nor Opus 5 has a verified entry on the Terminal-Bench leaderboard, and when the benchmark team verified Meta’s previous release, Muse Spark 1.1 landed at 76.2 against Meta’s published 80.0.
Sitting 3.8 points behind the leader while charging a fraction of the price is a coherent position, and Wang made it explicitly, saying the tool “can be an incredibly good option, especially from a cost perspective.” What that argument understates is that the standard tier already wins on cost. Claude Opus 5 lists at $5 and $25; Muse Code’s no-training tier undercuts it fourfold on input and nearly sixfold on output. Nobody has to hand over anything to get a cheap, near-frontier coding agent out of this launch.
That is roughly the outcome you would expect after four labs shipped the same agent inside a fortnight back in June. With the shape of the product commoditized, competition moves to model quality, price, and integration, and Meta led with price. Muse Spark 1.2 is also reachable through OpenRouter, so the model can be evaluated without adopting the harness or the billing arrangement, which is the sane way to try it.
The price is now public
The number worth following over the next few quarters is whether $0.10 per million tokens settles in as the going rate for training rights on working code. Meta has put that figure somewhere every rival lab and every enterprise procurement team can read it, and it will be hard to move up from there.
For a small agency running client work, the arithmetic resolves in the other direction. A few thousand dollars a year is a real saving, but Meta has not documented the scope of what you would be granting, and unlike a storage policy the grant cannot be walked back once the training run is done. Pay the $150 and keep the client’s code out of it.
Sources
- Meet Muse Spark 1.2 and Muse Code: a coding model and the agent built to run it — Meta AI Developers blog
- Meta debuts first AI coding agent to take on Anthropic and OpenAI — CNBC
- Meta launches Muse Code, an AI agent for large code bases — TechCrunch
- Meta wants to get inside your terminal with its new coding agent — The Register
- Meta’s New Mac Coding Agent Costs Up to 20x Less If You Let Meta Train on Your Data — MacRumors
- Meta Prices Muse Code 21x Cheaper for Developers Who Share Data — Implicator.ai
- Meta AI Releases Muse Code (Beta): A Terminal Coding Agent Powered by the New Muse Spark 1.2 Model — MarkTechPost
- Muse Code Benchmarks (Aug 2026): Meta’s 82.9% vs Verified Scores — Kingy AI