What Meta released
Muse Ember is a Meta model, sold on Meta's own API and resold by NanoGPT and Vercel AI Gateway. Its model page says plainly what the lab built it for:
Muse Ember is our smallest multimodal model built for high-volume agentic workflow execution, from coding subtasks and tool calls to document summariz...Access is the other thing Meta spells out. That matters for anyone weighing a direct account with the lab against a reseller they already use:
Muse Ember is on Model API and in Muse Code, on pay-as-you-go and with any subscription.
Where to buy Muse Ember
The usual reason to compare sellers of one model is that they price it differently. Here they do not: Muse Ember is billed $0.10 for Input and $0.50 for Output on every row of the table, which gives a mixed price of $0.20.
That mixed price weighs three input tokens for one output token, the usual shape of a chat exchange, so a single figure can stand for a whole price list.
| Seller | Input, $/M | Output, $/M | Mixed 3:1, $/M |
|---|---|---|---|
| NanoGPT | 0.10 | 0.50 | 0.20 |
| Vercel AI Gateway | 0.10 | 0.50 | 0.20 |
| Meta | 0.10 | 0.50 | 0.20 |
The three sellers share the lowest price, to the cent. The match extends to the cache read, meaning Input the seller has already seen and bills lower: Muse Ember's cache read is $0.004 per million tokens on all three price lists. The sources do not say why the resellers copy the lab's rates exactly.
What does an agent task cost?
A listed rate per million tokens leaves out the one thing a bill depends on, which is how much the model writes. So take a single agent step: a long context of 120,000 tokens re-read from the cache, a new request of 30,000 tokens, and a written answer of 15,000 tokens.
| Seller | One task, $ | 1,000 tasks, $ | Output share, % |
|---|---|---|---|
| NanoGPT | 0.011 | 10.98 | 68 |
| Vercel AI Gateway | 0.011 | 10.98 | 68 |
| Meta | 0.011 | 10.98 | 68 |
At the rates all three sellers share, that step costs $0.011, and 1,000 of them come to $10.98. Most of that money goes to Output: the answer is the smallest of the three token counts, yet Output makes up 68% of the cost.
The cache explains that imbalance. At $0.004 per million, the re-read context barely registers, while each Output token is billed at $0.50 per million. For a team running Muse Ember at volume, the length of the answers is the line to watch, far more than the size of the context.
Who should consider it?
Meta's documentation narrows the target to the work that sits inside larger agent systems:
It runs the sub-agent tasks, tool calls, and routing inside larger agent workflows, and processes large collections of images, files, and records.That profile fits the cost structure of the agent step: workloads that re-read a large context and write short answers are where Muse Ember's price list is kindest. A team already buying it from NanoGPT, Vercel AI Gateway or Meta has no price reason to switch, since the rates match line for line.
Is the full context available everywhere?
Yes. NanoGPT, Vercel AI Gateway and Meta each serve Muse Ember with a context window of 1,048,576 tokens, so an agent that carries a very long context can run on any of the three without losing room.
On InferIndex: Muse Ember model page