Muse Code Models, Pricing and Privacy: What to Check

Choose an explicit model and data tier

For a Standard session, launch the model explicitly. The official model catalog lists Muse Spark 1.3 and its Contributor variant. The suffix is meaningful: a model selection can change the data terms that apply to your prompts and responses.

muse --model muse-spark-1.3 --reasoning-effort medium

For a team project, record the selected model alongside the task result. If a request fails because of availability or quota, resolve that issue before changing tiers. A cheaper model is not an interchangeable fallback when its data terms differ. Ask the project owner to make that choice explicitly.

Do not submit sensitive, personal or confidential material to Contributor, including code you must keep confidential. Check regional eligibility as well as training consent. These restrictions are in Meta’s terms, section 6.2.

API token prices and a worked example

The published API rates checked on September 22, 2026 are in US dollars per one million tokens. These are metered API rates, not monthly subscription prices.

Suppose a small evaluation consumes 200,000 uncached input tokens and 20,000 output tokens. Standard token charges would be 0.2 × 1.25 + 0.02 × 4.25 = $0.335. Contributor would be 0.2 × 0.10 + 0.02 × 0.20 = $0.024. This is arithmetic for that stated workload, not a prediction of the cost of fixing one bug.

An agent can call the model repeatedly while inspecting files and testing an edit. Measure the whole completed task, including unsuccessful attempts, before estimating a weekly budget. The example excludes separate tool charges, taxes and any account-specific adjustments. Cached input only belongs in the calculation when usage actually reports it.

Subscription access is tied to its Muse credential

Meta documents Everyday Usage, High Usage and Power Usage plans. Availability and benefits can differ by region, so use the price and allowance shown during your own onboarding. The subscription documentation does not establish a single monthly price for every reader.

Subscription access uses the credential connected during Muse Code onboarding. Additional API keys created in the account use pay-as-you-go billing. Before comparing invoices, identify which credential path each task used. Keep a short record of the date, selected model and whether the task ran through subscribed Muse access or a separate API key.

An API key in the environment takes precedence over a browser session. Check authentication priority if you signed in but usage appears under an unexpected account. Do not print a credential to prove which route is active; inspect your shell or CI configuration and the account’s billing view.

No training does not mean zero retention

Standard excludes training on prompts and completions. It does not by itself establish zero retention. Meta’s Model API terms describe retention and other processing, including service operation and safety. Read those terms for the selected tier before sending proprietary code.

Treat three questions separately: may the provider train on this content, how is submitted content retained, and what do local tools store? A local log setting cannot answer the provider-retention question. Likewise, deleting a local transcript is not evidence that a hosted service deleted the requests it received.

For a first evaluation, use a repository you are authorized to submit. Avoid production credentials and customer records in sample prompts. If a result needs an error log, reduce it to the lines that reproduce the problem and remove secrets before giving it to the agent. This also makes the technical task easier to inspect.

Compare models with the same acceptance condition

Find the cause of this failing validation case.
Explain the relevant call path before editing.
Make one correction and run the existing targeted test.
Report the diff, test result and any unresolved assumption.

Run a representative task on a clean checkout and save its starting commit. Compare the correctness of the edit, review time and total usage. A fast answer that introduces a second defect is a poor result even if the token bill is small. Repeat a comparison only when the task or configuration changes enough to justify it.

Keep model choice stable while diagnosing Windows setup or adding MCP and skills. Changing the model, credentials and tools together makes it harder to identify why the result changed.

What CodeAgentSwarm’s beta establishes

Use the first-task guide to validate the standalone CLI before adding a desktop integration. Check CodeAgentSwarm’s public release notes for Muse availability.