Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Meta’s Muse Spark 1.3 Contributor tier is the cheapest listed option for input and output tokens, at $0.10 per million input tokens and $0.20 per million output tokens. The discount has a condition written into Meta’s own pricing page: you grant Meta permission to use your prompts and completions to train future Meta models. For a coding agent that reads repositories, diffs and test logs, “prompts and completions” can cover a large amount of source code. Whether the trade makes sense depends on what the work contains and on whether you are allowed to make it.
The current rate card, line by line
Meta’s Model API pricing page, retrieved October 7, 2026, lists two tiers for Muse Spark 1.3. Cached input is priced separately from uncached input, so the comparison needs three rows, not one.
| Line item (per million tokens) | Muse Spark 1.3 Standard | Muse Spark 1.3 Contributor | Standard price as a multiple of Contributor |
|---|---|---|---|
| Input (uncached) | $1.25 | $0.10 | 12.5x |
| Cached input | $0.15 | $0.002 | 75x |
| Output | $4.25 | $0.20 | about 21x |
The multiples are simple division of the listed rates. Guides that describe the Contributor tier as “20x cheaper” are rounding. That shorthand fits output tokens reasonably well, but it understates the uncached input gap and overstates the cached-input gap. Compare the line item that matches your workload, and use the current page rather than older guides, which were written when Muse Spark 1.2 was the current model.
What the discount exchanges for
The Contributor tier’s description on Meta’s pricing page reads: “Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models.” That sentence is the core of the trade. The lower price is conditional on the training permission, and the permission covers prompts and completions, not only the metadata around them.
#1 Best Overall
The Standard tier is described differently. Meta says Standard-tier prompts and completions are not used to train Meta models. That is a statement about the published tier distinction. It is not a general promise about every Meta product, how long data is retained, or how it is processed elsewhere. Meta’s model listing uses shorter labels, “Used to improve our products” for Contributor and “Not used to improve our products” for Standard, but the pricing page is the more precise source for the training wording.
What a coding agent actually sends
Meta announced the Muse Code beta on August 5, 2026, through Meta AI Research. It is a terminal coding agent that was powered at launch by Muse Spark 1.2, and Meta says it handles software-engineering work across repositories, including planning, writing code and validating results. Agents like this do not send only the question you typed. Depending on the harness and the task, the content of a model request can include more than you would paste into a chat window.
Three kinds of material are the ones to watch. Each is an example of what a harness may send, not a fixed list for every tool.
Rank #2
Repository source and diffs
A coding agent often needs to read the files it is changing, plus surrounding modules, so that its edits fit the codebase. Proposed changes and diffs may then be part of the request. If the source is proprietary, a diff is as sensitive as the file it came from.
Test output and error logs
When the agent runs a test suite and reports the result, the output may include stack traces, internal hostnames, fixture data or sample records. These are easy to overlook because they look like noise, but they often contain the most sensitive detail in the session.
Files the agent reads on its own
An agent may open files you never mentioned in the prompt, such as configuration, environment templates or documentation. If the harness forwards what the agent reads, those files can enter the request even though you did not choose to share them.
A third-party setup guide from Cobalt describes this flow for its own Muse Code task setup. It states that traffic can include repository source, diffs, test output and files read by the agent. That is one harness’s documented behaviour. Other integrations may send less, send it differently, or send it through a different route. The Meta material reviewed for this article does not specify which tier a Muse Code session bills against, so check your account and integration settings before assuming either tier applies.
A worked cost example, with its assumptions stated
The figures below are arithmetic on the listed rates. They are not a benchmark, and they are not a forecast for any real team. The assumptions are 10 million uncached input tokens and 1 million output tokens in a month, with cached input excluded and no other charges counted.
| Item | Standard | Contributor |
|---|---|---|
| 10M input tokens | $12.50 | $1.00 |
| 1M output tokens | $4.25 | $0.20 |
| Total for this assumed volume | $16.75 | $1.20 |
At this volume the difference is about $15.55 a month. A workload that is mostly cached input, or one that generates long outputs, will show a different gap. Run the numbers with your own token counts before you decide.
Rank #4
A decision framework for choosing a tier
Price should be the last question, not the first. These steps follow the documented tier distinction and the possible inputs described above. They are editorial recommendations, not additional Meta policy.
- Classify what the agent can see. List the repositories, branches, logs and config files that the session can read, including files it might open on its own.
- Sort the material into approved and not approved. Public open-source code you are free to share is a different category from proprietary source, customer records, personal data and credentials.
- Check the training permission against your obligations. The Contributor tier grants Meta permission to train on prompts and completions. If an employment contract, client agreement or regulation forbids that, the discounted tier is not an option for that work.
- Choose the tier per workload, not per company. A team can use Contributor for approved open-source work and Standard for everything else, provided the integration can route requests to the correct model ID.
- Keep secrets out of the working set. Credentials, tokens and production data should not sit in directories the agent can read, whichever tier you choose.
- Have security or legal review the actual integration. The harness, the provider terms and your own data-handling rules together determine the outcome. A rate card cannot settle that alone.
What to confirm before you switch
- The exact model ID. Muse Spark 1.2 and 1.3 are listed as separate model IDs with their own tier rates.
- The current rates on Meta’s Model API pricing page, including the cached-input line, which is the one most likely to change your bill.
- The context window for the model you select. Meta’s product page lists 1 million tokens for Muse Spark 1.3.
- Whether your harness forwards files the agent reads, and whether your tool shows the exact request contents.
- Whether the terms that apply to your account match the public pricing page you are reading.
Rates, model versions and terms can change. The figures in this article reflect Meta’s pricing page as retrieved on October 7, 2026, and should be checked again at the time of use.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
