Meta released Muse Spark Open on October 15, 2026. Downloadable weights give teams another deployment option. They do not make inference free or establish that running the model yourself will cost less.
As checked on October 16, Artificial Analysis lists hosted pricing of $1.25 per million input tokens and $4.25 per million output tokens. The relevant comparison is what a completed, acceptable task costs at those rates, including the model's actual token usage.
The $0.84 figure is an API benchmark result
Artificial Analysis reports an average Intelligence Index task cost of $0.84 at its xhigh configuration. That is a result for its benchmark, configuration and hosted pricing. It is not a self-hosted cost estimate, a price per arbitrary business task or a guarantee about your coding workload.
flowchart TD
A["Intelligence Index benchmark and xhigh configuration"] --> B["Measured token usage"]
B --> C["Hosted API pricing"]
C --> D["Reported average: 0.84 USD per benchmark task"]
D --> E["Run your own tasks before estimating savings"]
A per-token comparison is a starting point. A model that uses more output or needs more retries can erase part of an apparent rate advantage. Compare success rates and total costs under equivalent acceptance criteria.
Self-hosting needs a separate calculation
The model's total parameter count and numerical precision determine weight storage; active parameters per token do not make the inactive weights disappear. For illustration, 816 billion parameters at two bytes each require 1.632 trillion bytes for weights alone. Quantization changes that figure, while runtime state, KV cache and serving overhead add to it.
There is no single “roughly 1TB” hardware requirement without specifying precision and the serving setup. Nor does a hardware purchase price alone establish a break-even point. Utilization, throughput, redundancy, power, engineering and support all belong in the calculation.
Teams may value control or deployment flexibility even when hosted APIs are cheaper. That is a decision about requirements, not proof of a universal saving.
Read the actual license before calling it open source
Open weights and open source are not interchangeable labels. Consult the license linked from Meta's release for the rights and restrictions that apply to your intended use. This article does not assert an unverified revenue threshold or a blanket permission to offer a hosted service.
Decide using one representative workload
Start with a bounded evaluation through an available provider. Record the model configuration, token usage, accepted results, failures and latency. Price the same workload under a concrete self-hosted design only if control, scale or other requirements justify that work.
Use the $0.84 benchmark as a question to investigate: can this model complete your tasks economically? Do not put it in a self-hosting budget as an answer.
