Muse Spark Open API pricing and cost calculator

How this is calculated

The calculator opens on a request of 100,000 input tokens and 5,000 output tokens, half the input cached, with 100 requests a month. One uncached call costs $0.14625. With half the input read from cache it's $0.09125, and the starting total is $9.125. A real coding session looks different from one call. Picture an agent working through a bug across 60 turns. Each turn resends about 80,000 tokens of files, tool results and history, nearly all of it read from cache, and produces 2,000 tokens of reasoning and edits. That session costs $1.758. If the cache never hits, because the prefix changes every turn or the cache expires between steps, the same session costs $6.51. The gap between those two numbers is the reason to keep the prompt prefix stable. Put system instructions and files first, append new messages last, and don't reorder tool results. A team that runs 40 sessions a day over 22 working days is $1,547.04 a month at the cached rate. Add cache writes, tool fees and failed attempts on top, then compare the total with the number of merged changes. Output deserves its own attention. Reasoning can run long, and a 20,000 token answer after a 100,000 token prompt costs $0.155 with half the input cached. Media works differently. Screenshots, screen recordings, audio notes and files all count as input, but the number of tokens each one takes depends on the provider, so send real samples and read the usage figures. Meta publishes the weights on Hugging Face as a gated download under a custom license. Read that license before commercial use, since this page doesn't state its terms. The prices shown are the maker's own. Open weights allow other hosts, and a lower host price will appear in the calculator when one is listed.

Verdict

Muse Spark Open fits teams that want an open weight model for coding agents and may host it later. Measure first. Log tokens per turn, cache hits and how many turns a task takes, then price whole sessions here. Cost per merged change is a better number than cost per call. The weights are gated and use a custom license, so read it before you commit to self hosting. At 816B total parameters, that means a multi GPU server, and the VRAM calculator page for this model shows the memory budget and how much the long context adds.