As of October 21, 2026, Muse Code costs $5 to $200 a month on a subscription, or $1.25 per million input tokens and $4.25 per million output tokens on pay-as-you-go. Meta's Muse Code product page also lists a Contributor tier at $0.10 and $0.20, and that tier lets Meta train on your code. Subscriptions start in Contributor mode, so check the setting before you point it at a private repo.
What happened: Meta added plans, a new model and Windows
Meta released Muse Code in beta on August 5, 2026, calling it "a terminal coding agent powered by Muse Spark 1.2." On September 2, Meta said Muse Spark 1.3 is "now available on Muse Code and Meta Model API." Meta's product page lists 1.3 at the same token prices as 1.2.
- What changed: four monthly plans, Muse Spark 1.3, Windows support and higher Contributor rate limits.
- Who's affected: developers and team leads choosing a plan, and anyone whose subscription is still in Contributor mode.
Before and after: August 5 vs October 21
Muse Code launched in beta; it now has subscriptions and Windows support.
| At launch (Aug 5) | Now (Oct 21) | |
|---|---|---|
| Model | Muse Spark 1.2 | Muse Spark 1.3 or 1.2 |
| Ways to pay | Not confirmed | $5, $15, $50 or $200 a month, or per token |
| Platforms | macOS, Linux (per Meta's 1.3 post) | macOS, Linux, Windows |
| Contributor rate limit | 60 RPM, 2.1M TPM (MacRumors) | 150 RPM, 3M TPM (Meta docs) |
Reports on the old Contributor limit disagree: MacRumors gave 60 RPM, while The Batch and eesel gave 100 RPM. On Windows, Meta's docs say PowerShell replaces Bash and that voice input and session messaging are unavailable.
The plans scale by usage: Everyday ($5) allows "10-50 prompts every 5 hours." High Usage ($15), Power Usage ($50) and Maximum Usage ($200) give 5x, 20x and 80x that. Meta publishes no token allowance, so you can't compare a plan with per-token billing directly.
Contributor mode trains on your prompts and code
Meta's pricing and rate limits page says Contributor mode runs "the same models at heavily discounted token prices" in exchange for training rights. Those rights cover "the prompts, code, files, and context you send, plus the model responses you get back." Standard mode costs more, and there "your content is not used to train Meta models."
Meta's subscription docs say "Muse Code subscriptions default to Contributor mode" and that "you can turn it off at any time." Meta says Contributor mode gives you more usage from your plan, so turning it off means fewer prompts per 5 hours. Pay-as-you-go keys pick a tier per model name, such as muse-spark-1.3-contributor.
The matrix above lists both modes' token prices, rate limits and training use, all from Meta's docs.
Contributor mode makes sense for open-source repos, side projects and throwaway prototypes, where training use costs you little. Keep client code, credentials, regulated data and unreleased product work on Standard, the same rule as in our guide to what not to paste into ChatGPT. One CTO told The New Stack, "I won't gamble with my IP."
Heavy users also lose capacity on Contributor: 150 requests per minute per team against 3,000 on Standard.
What one agent step costs on each tier
On a step of 60,000 input tokens and 3,000 output tokens, Standard costs $87.75 per 1,000 steps with no cached input (cold) and $21.75 with cached input. Contributor costs $6.60 cold and $0.72 cached. These are our arithmetic from Meta's listed prices, not figures Meta publishes.
The discount depends on the token type. Contributor input is 12.5x cheaper ($1.25 vs $0.10), output is 21.25x cheaper ($4.25 vs $0.20), and a cold step works out 13.3x cheaper. So the discount is 12.5x on input and 21.25x on output; a headline "20x cheaper" figure only holds for output tokens.
What to do now
- Today: open your Muse Code subscription settings and turn Contributor mode off for any account that touches private code; expect less usage per plan.
- This week: split repos into public and private, and give each its own API key and model name.
- Before a big run: price it with the per-step figures above, and add $2.50 per 1,000 web search queries if you use grounding.
- If you compare agents: read our Claude Code capability guide for the alternative most teams weigh.
Still unconfirmed: default model and zero data retention
Meta's Muse Code docs still say "The default model is muse-spark-1.2," while the product page lists 1.3 first. Meta's 1.3 launch post also still says to install "on macOS or Linux." On zero data retention, Implicator reported it is "a separate option that Meta has only started accepting requests for," and Meta's pricing page mentions it only for Muse Voice Transcribe.
eesel, which sells a rival product, reports that reasoning tokens bill as output and that there is "no spend cap." Meta hasn't confirmed either. Model details are on our Muse Spark 1.3 and Muse Spark 1.2 pages.
FAQ
Is Muse Code free? No. Plans start at $5 a month, and the Contributor tier is discounted, not free.
Does Muse Code train on my code? Only in Contributor mode, which is the default for subscriptions. Standard mode content is not used for training.
Does Muse Code run on Windows? Yes, according to Meta's docs, without voice input or session messaging.
Is Muse Spark 1.3 better than 1.2? Meta says its engineers saw "~20% fewer tool calls and ~25% fewer tokens." That is Meta's own comparison, not an independent test.
