# Muse Spark 1.3 Charges 21× More for Output It Won't Train On

**Summary:** Muse Spark 1.3 charges $4.25 per million output tokens when prompts and completions stay out of training, versus $0.20 on Contributor. I see a data-classification choice, not a default bargain.

- Canonical: https://markhuang.ai/news/muse-spark-1-3-20-cent-output-data-clause
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-09-02
- Section: News
- Tags: Muse Spark 1.3, Meta Model API, AI Coding, AI Pricing, Data Privacy
- Source: [Meta for Developers](https://developer.meta.com/ai/models/muse-spark/)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![A stream of luminous data blocks splits between an open glass reservoir and a sealed black vault](https://cdn.markhuang.ai/news/muse-spark-1-3-20-cent-output-data-clause/hero.webp)

*Muse Spark 1.3 offers two routes to the same model family. The cheaper route opens prompts and completions to future training; the standard route keeps them out.*

[Meta's Muse Spark 1.3 page](https://developer.meta.com/ai/models/muse-spark/) pitches a stronger agentic and coding model with a 1 million-token context window. I kept looking at the pricing table below it. The Contributor tier charges $0.10 per million input tokens and $0.20 per million output tokens, while Standard charges $1.25 and $4.25.

The discount changes the product decision. Meta's [pricing documentation](https://dev.meta.ai/docs/pricing-rate-limits) says prompts and completions on Standard are not used to train Meta models. Contributor grants Meta permission to use both for future model training. To me, these are two data policies that happen to serve the same model. They do not belong in a dropdown as if price were the only difference.

The gap is hard to shrug off. Standard input costs 12.5 times as much, output costs 21.25 times as much, and cached input costs 75 times as much. Contributor can save real money. Whether a workload belongs there is a separate question.

## A better agent does not remove the deployment choice

Meta [released Muse Spark 1.3 on September 2, 2026](https://research.meta.ai/blog/introducing-muse-spark-1-3) for Muse Code and Meta Model API. It says the model handles long tasks and conflicting context better, asks for help when stuck, and confirms consequential actions. In comparisons run by Meta engineers, it used about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2.

Every unnecessary turn adds latency and cost, so those improvements could matter. The official benchmark graphic also reports 75.4 on DeepSWE v1.1 and 88.8 on Terminal-Bench 2.1 for Muse Spark 1.3 at max reasoning. I would not buy on those two scores. They come from Meta's selected comparison, and the launch post says max reasoning will arrive only after additional safety testing. The strongest mode in the chart was not the fully available launch-day product.

Independent testing is still one version behind. [Artificial Analysis measured Muse Spark 1.2](https://artificialanalysis.ai/articles/muse-spark-1-2) at $0.40 per task on its Intelligence Index, but it also found about 53% more input tokens and 36% more output tokens per task than 1.1. A token price is not a task price. I made the same case in [my recent look at token sticker prices](https://markhuang.ai/news/token-sticker-price-is-a-trap): the useful unit is a result that passes review.

## The discount is permission, not a lesser model

Meta lists the same 1,048,576-token context window for the Standard and Contributor model IDs. The published distinction is what Meta may do with the traffic.

| Endpoint                     | Training use                                              | Input per 1M | Output per 1M |
| ---------------------------- | --------------------------------------------------------- | ------------ | ------------- |
| `muse-spark-1.3`             | Prompts and completions are not used to train Meta models | $1.25        | $4.25         |
| `muse-spark-1.3-contributor` | Prompts and completions may train future Meta models      | $0.10        | $0.20         |

I can see a good use for the Contributor tier. Public open-source repositories, synthetic test cases, benchmark harnesses, and throwaway prototypes can generate plenty of useful traffic without exposing material the team needs to protect. The low rate makes broad experiments much easier to justify.

Private repositories, customer records, incident logs, unreleased product plans, and any prompt that might capture a credential stay on Standard for me. Coding agents gather context aggressively. A request that begins with a harmless bug description can pull in source files, terminal output, configuration, and issue history before it is done. The endpoint decision covers the whole agent loop, far beyond the sentence a developer typed.

> **Info:**
>
> My rule would be simple: use Contributor only when I would be comfortable treating the complete prompt-and-completion trace as training-eligible data. If that answer changes halfway through the task, the workload started on the wrong endpoint.

## The missing details matter before production

The pricing page clearly states the exchange, which is better than hiding it in a footnote. It does not, on that page, give a retention period, a deletion workflow, or a more granular account of how training-eligible records are reviewed and prepared. I would want those answers before approving Contributor traffic from a company repository.

The wording also makes opt-in a model identifier. That is easy to automate, but easy to misconfigure. A copied environment variable or fallback route could move a sensitive job onto the discounted endpoint without changing the code that assembles its context. Cost dashboards would celebrate while the data boundary quietly moved.

I would separate the tiers at the infrastructure layer with different project credentials and budget labels. Contributor workloads would need an explicit allowlist. I would also log which model ID handled each job. Redaction can help, but it should not carry the whole policy. An agent can reveal sensitive structure even after obvious secrets have been removed.

## The price is good when the data fits

Muse Spark 1.3 looks competitive enough to test, and Meta's reported efficiency gains make that test more interesting. The Contributor price makes large experiments unusually cheap. Neither point turns training permission into a minor billing detail.

I would use the cheap tier deliberately for public and synthetic work, then pay Standard rates when the agent can touch anything I would not publish. That sounds conservative until the 21.25-times output gap reaches the budget meeting. Then it becomes exactly the decision the pricing table was designed to force: how much is the organization willing to save, and which data is it willing to contribute?