Meta Muse and Llama API pricing, newest model first

Muse rows show Meta's price on its own API, in dollars per million tokens. One of them is a trade. In Contributor mode, Meta can use your prompts, files, code and the answers it sends back to train future models.

In return the rate drops. Muse Spark 1.3 Contributor costs $0.10 input and $0.20 output per million, against $1.25 and $4.25 in the normal mode. That's a good deal for public data, test runs and hobby projects. Keep customer data, private code and anything under contract on the normal rate.

A token read back from the prompt cache costs $0.004 per million on Muse Ember, 96% off the normal input rate. Keep the start of each prompt the same so it gets read back.

Llama is still around, and it's open weight, so many hosts run it. Llama 4 Maverick shows Parasail, at $0.35 input and $1 output per million, and you can also run it on your own hardware. Muse has open weight models too. Muse Spark Open is a coding model you can download, and Muse Glimmer 30B is a dense model distilled from Spark that fits on consumer hardware.

Llama 4 Scout has the largest window, 1.31M tokens, at the host shown, DeepInfra. Other hosts can cap it lower, so check before you send a huge prompt.

Use Muse Ember for routing, tool calls and other small steps, and Muse Spark 1.3 for the long agent and coding work. Keep Muse Wonder for the hard reasoning problems, and track the cost per accepted answer as you go. Only use Contributor mode for data you'd be fine seeing in a training set.

If you'd rather run a model yourself, look at the Llama models and the open Muse releases. You pay for GPUs instead of tokens, and nothing leaves your own servers.