Muse Spark 1.3 Pricing: How Meta Undercuts GPT-5.6 and Opus 5
Meta prices Muse Spark 1.3 at $1.25 per million input tokens and $4.25 per million output tokens. Here's how that compares to GPT-5.6 and Opus 5.
What does Muse Spark 1.3 cost per token?
Meta has priced Muse Spark 1.3 at $1.25 per million input tokens and $4.25 per million output tokens. That places it well below what frontier labs typically charge for models that claim comparable agentic and coding performance, and it’s a notable move for a model Meta says it plans to open weight soon. The pitch is simple: benchmark scores that go toe-to-toe with GPT-5.6 and Opus 5, at a fraction of the metered cost.
TL;DR
- Muse Spark 1.3 is priced at $1.25 per million input tokens and $4.25 per million output tokens, undercutting rival frontier pricing by a wide margin.
- Meta is positioning the model as a multimodal reasoning system built for long horizon agentic work, coding, and computer use, with a full 1 million token context window.
- On benchmark charts, Muse Spark reportedly tops professional tool use and agentic computer use, edging out both GPT-5.6 and Opus 5 in those categories.
- The model’s standout result is on long context retrieval, scoring in the high 90s on a million token test while competitors fade into the 60s and 70s.
- Coding benchmarks show a tight three way race between Muse Spark, GPT, and Opus, with no single model running away with it.
- A real world test involving an independent AWS deployment task (provisioning CloudFront and S3 with no manual steps) cost around $2.49 in total API spend across several tests, according to hands on testing.
- Mark Zuckerberg has confirmed Meta intends to release Muse Spark 1.3 as open weight in the near future, which changes the calculus for anyone comparing hosted pricing against self-hosting costs.
How does Muse Spark 1.3’s pricing compare to GPT-5.6 and Opus 5?
Meta’s per-token pricing for Muse Spark 1.3 sits at $1.25 input and $4.25 output per million tokens. Independent testing referenced benchmark charts where Muse Spark matches or beats GPT-5.6 and Opus 5 on several agentic tasks, while reportedly costing meaningfully less to run. Exact current pricing for GPT-5.6 and Opus 5 wasn’t detailed in the source material, but the framing throughout was consistent: Meta is deliberately pricing this model “almost too cheap to meter” relative to what the other two labs charge for similar capability tiers.
For teams running high-volume agentic workloads (think autonomous coding agents, long-running research tasks, or multi-step tool use pipelines), the gap between $1.25/$4.25 per million tokens and pricier frontier alternatives adds up fast. A workload that burns through millions of tokens a day in tool calls, retries, and long context retrieval turns a per-token discount into a real budget difference by the end of the month.
What can Muse Spark 1.3 actually do?
Muse Spark 1.3 is built as a multimodal reasoning model aimed at long horizon agentic work: coding, computer use, and tasks that require staying coherent across many steps without losing track of the original goal. It ships with a full 1 million token context window, which matters directly for its pricing story. A cheap per-token rate is only useful if the model can hold enough context to do something useful with it, and Meta’s own benchmark data (per the model’s public benchmark chart) shows Muse Spark scoring in the high 90s on a million token retrieval test, while rival models drop into the 60s and 70s, with Opus reportedly posting no score at all on that specific test.
On agentic benchmarks, Muse Spark reportedly leads on professional tool use and agentic computer use, ahead of both GPT-5.6 and Opus 5. It’s described as roughly even with Opus 5 on harder end-to-end business workflows. Coding is the closest contest, with all three models clustered tightly together on long horizon agentic coding tasks. Opus 5 reportedly still leads on knowledge work and some agentic categories, so this isn’t a clean sweep for Meta. It reads as a competitive, honest benchmark spread rather than a one-sided win, with Meta’s real edge being consistency over long task horizons.
Does the pricing hold up in real-world testing?
One practical test involved handing the model a live AWS account with no existing resources and a single instruction: build a self-contained animated website and deploy it globally using S3 for storage and CloudFront for content delivery, with no step-by-step guidance. The model had to plan and execute every command itself, provisioning the CloudFront distribution, wiring it to an S3 bucket, generating the site’s HTML and animation logic, and returning a working URL.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
The run succeeded end-to-end: the model created the S3 bucket, generated the HTML file with an animated scene, tied the CloudFront distribution to the correct S3 origin, and returned a live, working URL. During execution, the model appeared to parallelize independent subtasks (provisioning infrastructure while simultaneously drafting site content) rather than working strictly sequentially, which is consistent with Meta’s claim about long horizon coherence. Server throttling was noted during testing, suggesting demand or load issues on Meta’s infrastructure shortly after launch.
Beyond the deployment task, the model was also tested on a vision and situational awareness task (correctly interpreting a WhatsApp conversation with an intentional double meaning), a combined chemistry and math reasoning problem (correctly working through a two-stage buffer titration and identifying the dominant reaction), and a multilingual task asking for culturally accurate drink names across roughly 79 languages, complete with correct native scripts across Arabic, Cyrillic, and Devanagari, among others. Across all of these tests combined, total API spend came out to roughly $2.49, according to the testing session referenced here. That figure is illustrative of one testing session rather than a benchmark of typical costs, but it’s a useful anchor point given how much ground the tests covered.
Is Muse Spark 1.3 worth it for developers?
For anyone building agentic tools, coding assistants, or long-context retrieval systems, the combination of low per-token pricing and a 1 million token context window is the headline. If the benchmark claims and real-world testing hold up broadly, Muse Spark 1.3 offers a way to run frontier-tier agentic workloads without frontier-tier bills. The fact that Meta is also planning to open weight the model soon adds another dimension: teams that want to self-host rather than pay per token at all will have that option once weights are released, which could make Muse Spark relevant both as a hosted API and as a deployable open model.
The caveats are real. Benchmarks are not a clean sweep, Opus 5 still leads in some categories, and early throttling suggests infrastructure is still catching up to demand. But on pricing alone, Muse Spark 1.3 is aggressive, and for cost-sensitive agentic workloads that lean on long context and tool use, it’s a serious option to evaluate against GPT-5.6 and Opus 5.
Frequently Asked Questions
How much does Muse Spark 1.3 cost per million tokens?
Meta prices it at $1.25 per million input tokens and $4.25 per million output tokens.
Is Muse Spark 1.3 open source?
Not yet at launch, but Meta founder Mark Zuckerberg has announced plans to release it as open weight soon.
How does Muse Spark 1.3 compare to GPT-5.6 and Opus 5 on benchmarks?
It reportedly leads on professional tool use, agentic computer use, and long context retrieval, is roughly even with Opus 5 on complex business workflows, and is tightly matched with both on coding. Opus 5 still leads in some knowledge work categories.
What is Muse Spark 1.3’s context window?
It supports a full 1 million token context window, which it demonstrated with a high-90s score on a million-token retrieval benchmark.
What kind of tasks was Muse Spark 1.3 tested on?
Documented tests include autonomous AWS deployment (provisioning S3 and CloudFront from a single prompt), reading and interpreting a multi-message WhatsApp conversation, solving a two-stage acid-base chemistry problem, and generating culturally accurate drink names across roughly 79 languages.
