Muse Spark pricing: $1.25 / $4.25, or ten cents if Meta may train
A price-comparison review of the Muse Spark 1.3 API, not the Muse app. Standard is $1.25 input and $4.25 output per million tokens. The contributor tier is much cheaper because Meta may train on your prompts and completions. The $20 and $100 plans are the consumer app.
| Bill | What Meta printed | What it is not |
|---|---|---|
| muse-spark-1.3, $ per 1M in / cached / out | $1.25 / $0.15 / $4.25 | Not used to train, per the card |
| muse-spark-1.3-contributor | $0.10 / $0.002 / $0.20 | Used to improve Meta’s products |
| Context, both ids | 1,048,576 tokens | Not a consumer-app quota |
| Muse app Power | $20 / month | More app usage, not this token rate |
| Muse app Maximum | $100 / month | More app usage, not this token rate |
| Reasoning tokens | Billed as output | A deeper effort is a larger bill |
| Web search grounding | $2.50 / 1,000 queries | On top of tokens, if the tool is on |
The API number, dated
Meta’s Muse Spark 1.3 card lists two checkpoints. Standard muse-spark-1.3 is $1.25 input, $0.15 cached input, $4.25 output per million tokens, and the card says this tier is not used to improve Meta’s products. Contributor muse-spark-1.3-contributor is $0.10 input, $0.002 cached input, $0.20 output, and the card says that tier is used to improve their products. Same 1,048,576-token context on both. Source: Muse Spark 1.3 model card.
That is the builder bill. It is not the consumer bill. The Muse app’s paid plans, Power at $20 a month and Maximum at $100 a month, buy more agent usage inside the US app. They do not buy these token rates. Free-tier caps for the app were not printed as a token allotment. Putting $20 next to $1.25 in one cell is how this product gets misquoted. The app story is on the desk note.
What the cheap tier is selling
Contributor is about a twelfth of standard input and about a twentieth of standard output. The discount is the right to train. If the prompts are customer data, source code, or anything you would not paste into a public training set, the $0.10 price is the expensive one. Meta’s own label is enough. You do not need a second interpretation.
Cached input on standard is $0.15 against $1.25 fresh. That matters for a repeated system prompt or a long file you keep sending. It does not apply itself. A client that never hits the cache still pays $1.25.
What the headline price leaves off the invoice
- Reasoning tokens are output. Meta’s developer notes say Muse Spark thinks before it answers, those tokens show up as reasoning tokens, and they are billed as output.
reasoning_effortruns from lighter settings up to max on the standard tier. Max is a quality knob and an invoice knob. - Agent runs are output-heavy. Tool loops, browser steps, and self-checks multiply the $4.25 side. We have not metered tokens per task on 1.3. A low input price does not cap that.
- Web search is a separate meter if you turn it on. Meta’s pricing notes list web-search grounding at $2.50 per 1,000 queries on Muse Spark text models, on top of tokens. Confirm the line still applies to 1.3 before you budget a browsing agent around it.
- The consumer checkout is a third bill. Stripe Link, one-time cards, and purchase protection are features of the Muse app. They are not discounts on the Model API.
Against the other list prices on this desk
Set next to numbers we have already dated, standard Spark output at $4.25 sits under Grok 4.7’s $6 and under Claude Sonnet 5’s $10, and far under the Fable 5.1 Max $50 column on xAI’s sheet. Spark input at $1.25 matches the GPT-5 workhorse figure this desk dated at $1.25 / $10, with a cheaper output than that $10. None of those models shares Muse Spark’s benchmark card. A cheaper token is not a win on GDPVal-AA, where Opus 5 still leads the cell Meta printed. Price is the reason to prototype. The profile is the reason to stay or leave.
When the low price is the wrong reason
Switching a production workload onto the contributor checkpoint to save money is a training-data decision you should write down, not a procurement shortcut. Switching off Opus because Spark’s token price is lower, without checking the GDPVal and OSWorld rows, repeats the Grok mistake of quoting a Max SKU’s price against a different job. Switching the office onto the Muse app because the API is $1.25 is a category error: the app is a US consumer agent with its own monthly plans.