On September 2, Meta shipped Muse Spark 1.3, the next model in the Muse line. Multimodal, one million tokens of context, text, image and video as input. On paper, a solid frontier model. But the pricing Meta ships alongside it is more interesting than the benchmarks.
There are two ways to use Muse Spark 1.3. The standard endpoint keeps your data private and costs $1.25 per million input tokens and $4.25 for output, measured at the highest reasoning tier. Cached input runs $0.15 per million. That’s a normal, market-rate price.
The Contributor tier
Then there’s the Contributor tier. It costs $0.10 input and $0.20 output per million tokens - up to 21x cheaper. Cached input drops from $0.15 to $0.002, a cut of roughly 75x, basically free. What you give in return: Meta trains on your prompts and the model’s responses.
It’s an honest trade, I’ll give them that. Meta hides nothing and puts the swap right on the table: you pay less money, you pay with your data instead. For plenty of workflows - prototypes, internal tools, non-sensitive bulk processing - that can be exactly right. For anything where confidential content flows through, it’s exactly wrong.
The frontier asterisk
VentureBeat points to a catch worth knowing. Meta’s best benchmark numbers for Muse Spark 1.3 come from a model variant developers can’t broadly use yet. It’s a pattern you see more often lately: a lab shows top-tier scores that aren’t actually available in practice. Frontier performance with an asterisk.
I find the Contributor tier pretty clever as an idea - and a little unsettling at the same time. The rule used to be: if you pay through the API, your data stays off-limits for training. Meta flips that and turns data into hard currency. If this catches on, “it’s basically free” quickly becomes “it costs you your data.” For developers, the takeaway is mostly this: look carefully at which endpoint you’re calling. The price gap is so large the temptation is real - and a wrongly ticked box gets expensive, just not in dollars.
Sources: VentureBeat: Meta says Muse Spark 1.3 has frontier performance, Cryptopolitan: Meta cuts Muse Spark prices up to 21x for developers who hand over their prompts, Artificial Analysis: Muse Spark 1.3