Meta Muse Ember: ten cents per million tokens, but only up to 200,000 tokens of context
What is Muse Ember and what does it cost? Muse Ember (muse-ember-1.0) is, according to Meta, its smallest multimodal model, built for large numbers of agent tasks. It has 1 million tokens of context and costs 0.10 dollars per million input tokens and 0.50 dollars per million output tokens. That price applies up to 200,000 tokens of context; above that you pay three times as much: 0.30 and 1.50 dollars. Ember is an API model; the developer page mentions no downloadable weights.
The figures in this article come from Meta's developer page, read on November 10, 2026. Earlier we wrote about active memory and focus in Meta Muse; Ember is the small sibling in the same family.
The price tiers
| Context | Cached input | Input | Output |
|---|---|---|---|
| Up to 200,000 tokens | $0.004 | $0.10 | $0.50 |
| Above 200,000 tokens | $0.012 | $0.30 | $1.50 |
Prices per million tokens. If you use Ember with the full 1 million tokens of context, you pay the higher rate. "A million tokens for ten cents" only holds below the 200,000 threshold. Prices of other models are in our overview models and prices.
Where you can use Ember
The page names three routes: the Meta Model API with your own account, a listing on Google Cloud Marketplace (muse-ember-1-0), and Muse Code subscriptions from 5 to 200 dollars a month. Whether Ember is included in every plan is not stated. The page says nothing about the maximum output length.
Benchmarks: Meta's own comparison
Meta sets Ember against Claude Haiku 5.5 and GPT-6 Luna, all three at their highest setting. According to Meta, Ember scores 61.6 on Agents' Last Exam (Haiku 5.5: 57.0) and 40.9 on Terminal-Bench 4.0 (Haiku 5.5: 39.2). On FrontierCode 2.0 Ember trails: 27.2 against 34.9 for Haiku 5.5. On the Artificial Analysis Index, Ember and Haiku 5.5 both sit at 36. These are figures Meta publishes; we did not measure them ourselves. Background on agents and the models behind them is in our overview Claude and agents.
A small inconsistency
The page calls Ember the "smallest multimodal model built for high-volume agentic workflow execution", while a callout on the same page speaks of "our smallest reasoning model". That does not change the price, but it shows the description is not settled yet.
What this means
If you run many small agent tasks, such as classification, summarization or subagents, Ember is one of the cheapest options this week, as long as context stays under 200,000 tokens. Self-hosting is not possible. More signals from that day are in the trending briefing of November 10.
What we did not check
We could not open Meta's blog post about Ember while writing, so we do not give a release date. The benchmarks and the partner quotes from You.com and Cloudflare come from Meta.
Checked on November 10, 2026.
This article was written with AI (Claude, Anthropic) and checked against the sources below.
Sources
- Meta, Muse Ember developer page: dev.meta.ai
