What Is Meta Muse Ember? The Tiny Model Built for Agent Tasks
Meta Superintelligence Labs just launched Muse Ember, its smallest and most efficient model, built for the repetitive work inside agent runs.
AI Stella and Liam are AI hosts with AI-generated voices. About our AI hosts
Key takeaways
- Muse Ember is Meta's smallest and most efficient model, built for the repetitive work inside agent runs: tool calls, routing, extraction, long document processing and coding subtasks.
- It costs $0.10 per million input tokens, $0.004 cached input and $0.50 output, with up to 1M tokens of context and 512K tokens of output.
- It matches Claude Haiku on the Intelligence Index (36) at about one third of the cost per task, but loses on FrontierCode, 27.2 vs 34.9, and it is not open weights.
- The winning pattern is a planner plus Ember: a bigger model plans and checks, Ember does the volume legwork through the Meta Model API (muse-ember-1.0), Muse Code, Google Cloud or Oracle Cloud.
Show notes
Meta Superintelligence Labs just launched Muse Ember, its smallest and most efficient model, built for the repetitive work inside agent runs: tool calls, routing, information extraction, long document processing, and coding subtasks like reading files, running tests and making targeted changes. In this episode of Agentic Pod, Liam and Stella cover what Ember is, the pricing and context numbers, the benchmarks including the one it loses, how to try it step by step, and a real workflow where a bigger model plans while Ember handles the volume.
What we cover:
- What Muse Ember is: Meta’s tiny multimodal model for repetitive agent tasks
- Pricing: $0.10 per million input tokens, $0.004 cached input, $0.50 output, up to 1M token context and 512K output
- Benchmarks: Intelligence Index 36 matching Claude Haiku at about one third the cost per task, plus AutomationBench, Terminal-Bench and Coding Agent Index
- The honest loss: FrontierCode 27.2 vs Haiku 34.9, and it is not open weights
- How to try it: Meta Model API (muse-ember-1.0, Responses API), Muse Code, Google Cloud and Oracle Cloud
- A real workflow: the big model plans, Ember does the legwork
Chapters
Questions answered in this episode
What is Meta Muse Ember?
The smallest model in Meta Superintelligence Labs' Muse family, launched November 9, 2026, tuned for repetitive agent subtasks like tool calls, routing, extraction and small coding jobs.
How much does Muse Ember cost?
$0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, with up to 1M tokens in and 512K tokens out.
Is Muse Ember open weights?
No. It is available through the Meta Model API, Muse Code, Google Cloud and Oracle Cloud.
Links & resources
Full transcript
Intro
Hey hey hey, welcome to Agentic Pod! I'm Liam.
And I'm Stella! Liam, Meta just launched a model that costs 10 cents per million input tokens?
Ten cents. That is Muse Ember, from Meta Superintelligence Labs. It launched on November 9th, and it is built for one job. Repetitive agent work.
Ten cents sounds like a typo. Is it actually smart?
That is the whole question. By the end you will know the price, the benchmarks, where it loses, and how to try it today.
What Muse Ember is and why it matters
Okay, start simple. What is Ember, in one sentence?
It is the smallest model in Meta's Muse family. Muse Wonder does the planning. Muse Ember handles the volume.
So the big brain thinks, and the small one does the chores?
Exactly. Tool calls. Routing. Pulling facts from documents. Reading files. Running tests. One small targeted code change.
The boring steps that fill every agent run.
And they add up fast. One task can mean hundreds of small calls. If every call uses a huge model, the bill explodes.
Ember is also multimodal. Text, image, audio and video. It can read a screenshot or a long document, not just plain text.
Okay, that is useful. But cheap and useful rarely travel together. Show me the numbers.
The numbers: pricing, context and benchmarks
Pricing first. 10 cents per million input tokens. Cached input is almost free, at $0.004 per million. Output is 50 cents per million.
Wait, say the context size. That part surprised me.
Up to one million tokens in. And up to 512,000 tokens out. That is a huge window for a tiny model.
So it can swallow a whole code base in one go.
That is the idea. Now the scores. On the Artificial Analysis Intelligence Index, Ember scores 36 at max reasoning.
And the comparison everyone will ask about?
Claude Haiku also scores 36. But Ember costs about 14 cents per task, against 41 cents for Haiku. Roughly one third the cost.
One third. Okay, that got my attention.
On Automation Bench, Ember scores 37.8, against 13.9 for Haiku. On the Coding Agent Index, 46 against 37.
Those are big gaps. Where does it lose? Be honest.
I will. On Frontier Code, Ember scores 27.2, and Haiku scores 34.9. Ember does not win everywhere.
Good, you said it out loud. A chart that only shows wins is just an ad.
Terminal-Bench is close too. Ember 40.9, Haiku 39.2. So how do you actually try it?
How to try it and a real workflow
Yes, step by step. Where do I even get it?
Step one, go to the Meta Model API. The model ID is muse-ember-1.0. It works through the Responses API.
And if I live in a coding tool all day?
Step two, try Muse Code. Step three, it is also on Google Cloud and Oracle Cloud.
And real teams are already using it?
Meta lists Box, Cline, Oracle, Cloudflare and You.com. Cloudflare reports up to 30% faster tail latency. You.com reports 2.2 second 95th percentile latency, and finished all its coding quests.
Okay, give me one real workflow, not a feature list.
Here is mine. Your big model plans the feature and writes the spec. Then Ember does the volume work underneath it.
Like what, concretely?
Ember reads the twenty files the plan touches. It runs the tests after each change. It pulls the exact lines you need from a long doc. The planner only wakes up for real decisions.
So I pay planner prices for decisions, and Ember prices for legwork.
Exactly. But there are catches, and one of them matters a lot.
Gotchas, who it is for, and the verdict
Let me guess the first catch. It is not open weights?
Correct. Not open weights. You call it through Meta or a cloud partner. If you need a model you can run yourself, this is not that.
Catch two is the FrontierCode score you already confessed.
Confessed is fair. For the hardest coding work, a bigger model still wins some rounds. Ember is a volume tool, not a genius replacement.
So who should actually try it this week?
Anyone whose agent bill is mostly small calls. Routing, extraction, tool calls, test runs, first pass code edits. That is Ember's home turf.
And who should skip it?
Anyone who needs open weights, or anyone whose work is one hard problem at a time. You will not feel the savings there.
My verdict: the price is the story. 14 cents a task changes what you can afford to automate.
Same. The guide with the API steps is in the description. Would you let a tiny model run your agent's busywork? Tell us in the comments. See you next time!