Muse Spark 1.3 is available through Meta Model API, with a 1M-token context window and a pricing decision that deserves attention. The standard model costs $1.25 per million input tokens and $4.25 per million output tokens. Contributor lowers those rates to $0.10 and $0.20, with a different condition on product-improvement use.
For developers evaluating a new coding or document workflow, that difference matters more than the headline discount alone. Here is how the options fit together, what Meta’s published results show, and what to test before adopting the model.
Pricing and availability checked September 16, 2026. This SeeAPI article covers external access options; SeeAPI integration has not been confirmed.
What is Muse Spark 1.3?
Meta released Muse Spark 1.3 on September 2, 2026. Its official launch announcement emphasizes stronger coding and agentic workflows: tasks in which a model plans, uses tools, and works through several steps toward a result.
Meta describes better handling of long instructions, conflicting information, and user interruptions. It also reports roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 in internal engineer comparisons. Those are vendor-reported observations, not performance measurements from SeeAPI.
The practical question is whether those improvements help your application complete useful work with fewer retries. A model that generates a good first answer can still struggle when it must preserve requirements across a lengthy sequence of actions.
Muse Spark 1.3 API pricing
Meta’s official API pricing table lists the following dollar rates per million tokens. Tokens are the units used to measure model input and generated output.
API model ID | Context window | Input | Cached input | Output |
|---|---|---|---|---|
| 1M | $1.25 | $0.15 | $4.25 |
| 1M | $0.10 | $0.002 | $0.20 |
For a workload using 1 million uncached input tokens and 200,000 output tokens, the listed token charges would be:
Standard: $1.25 + $0.85 = $2.10.
Contributor: $0.10 + $0.04 = $0.14.
That is about a 93.3% reduction for this particular mix of input and output. It is an illustrative calculation, not a quote for a complete application or another provider’s bill.
Cached-input rates can reduce the cost of eligible reused input. They should not be applied to every request automatically. When budgeting, separate uncached input, cached input, and output, then measure how many requests a completed task actually requires.
Muse Spark 1.3 Contributor: understand the condition
The important distinction appears directly beside the model IDs in Meta’s pricing table. Contributor is marked as used to improve Meta’s products; the standard model is marked as not used for that purpose.
This makes Contributor worth evaluating for public or synthetic test material when that condition fits your intended use. Before sending private repositories, customer documents, or internal conversations, check the applicable terms and your organization’s requirements.
The standard listing’s product-improvement statement should not be expanded into an unsupported promise about retention or every aspect of data handling. Choose the tier on its documented conditions, then assess cost. Both listings advertise the same 1M context size; the cheaper rate does not mean a smaller advertised context window.
What does the 1M context window mean?
A context window determines how much information a model can work with in a request. A larger window can make it easier to include extensive documents, repository material, or accumulated task history.
However, capacity and reliable use of that information are separate questions. Loading more text can increase input spending without improving the answer. The useful test is whether the model finds the relevant evidence and retains constraints as the input grows.
For an initial evaluation, try a document bundle with known answers or a repository task with clear acceptance criteria. Increase the supplied context gradually and compare correctness, response time, and token consumption.
Muse Spark 1.3 benchmarks: promising, with qualifications
Meta’s model page publishes these selected results for Muse Spark 1.3 at max reasoning:
Benchmark | Meta-reported score | Evaluation focus |
|---|---|---|
MRCR 256K–512K | 98.5 | Long context |
MRCR 512K–1M | 98.1 | Long context |
DeepSWE v1.1 | 75.4 | Long-horizon agentic coding |
SWEAtlas CodeBase QnA | 59.4 | Codebase understanding |
Terminal-Bench 2.1 | 88.8 | Agentic terminal coding |
These are separate benchmark scores, not interchangeable measures of accuracy. They also describe a particular reasoning setting. Meta’s table lists the earlier 1.2 model at xhigh, so a version comparison would also involve different reasoning settings.
The results justify testing long-context and coding workloads. They do not establish your application’s latency, cost per successful task, or reliability with its own tools and instructions.
API access, OpenRouter, and Muse Code
Meta Model API offers direct, self-serve access and currently labels the service public preview. Meta describes compatibility with existing OpenAI SDK code, but developers should still check current documentation for supported request fields and features.
Meta also identifies OpenRouter as an access route for Muse Spark. One detail needs care: the OpenRouter button on the inspected 1.3 model page points to a 1.2 URL. Verify the exact 1.3 listing and provider terms before configuring that route; the button alone does not establish version-specific availability.
Muse Code is Meta’s terminal coding agent. Use this distinction when choosing how to experiment: the API connects a model to your application, while Muse Code supplies a coding workflow around it.
Muse Code’s listed subscriptions are $5, $15, and $50 per month for Everyday, High, and Power Usage. These are subscription plans with usage allowances, separate from the per-token API pricing above.
Should you test Muse Spark 1.3?
Yes—Muse Spark 1.3 merits a bounded evaluation for coding and large-context tasks. Start with representative inputs, a fixed spending limit, and measurable success criteria. Compare completed-task cost, retries, latency, and instruction adherence before expanding usage.
Contributor offers substantial potential savings when its data-use condition fits. Standard offers a different product-improvement condition at higher token rates. Choose deliberately, verify the access route, and treat production adoption as a decision supported by your own results.
