Images in this article are AI-generated concept illustrations, not screenshots from Meta's products.
Meta has just launched the smallest AI model it has ever built — and it is punching far above its weight. Muse Ember, released on 9 November 2026 by Meta's Superintelligence Lab, is designed for the repetitive work AI agents do all day: calling tools, running coding subtasks, routing requests and pulling data out of long documents. And according to independent benchmarks, it matches Anthropic's Claude Haiku 5.5 while costing about a third as much per task.
The price is the headline. Ember costs $0.10 per million input tokens — a fraction of what frontier models charge. Meta's pitch is simple: most agent work does not need a giant model, it needs a cheap one that does the boring jobs well, thousands of times a day.
Key facts: Launched 9 November 2026 · Smallest and most efficient model in Meta's Muse family · $0.10 per million input tokens, $0.50 per million output tokens · 1M-token context window, up to 512K output tokens · Artificial Analysis Intelligence Index score 36 — matching Claude Haiku 5.5 at about a third of the cost per task · Available now on the Meta Model API and Muse Code as muse-ember-1.0.
What Is Muse Ember?
Muse Ember is the fourth model in Meta's Muse series, and the smallest. It follows the mid-weight Spark (April 2026), its open-source twin Glimmer (August 2026) and the frontier Wonder model (October 2026). Where Wonder is built to be the smartest model Meta has, Ember is built to be the cheapest one that is still good enough — the workhorse for high-volume agent workflows.
Meta describes Ember as "designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." In practice, that means the model sits behind the scenes: an AI agent calls Ember dozens or hundreds of times to complete one job, and the cost of each call is what decides whether that product is affordable to run.
Ember is also the first model Meta trained using Muse Wonder Full, its biggest model — a case of the flagship building its own smaller sibling. It accepts text, images, audio, video and PDF documents, and returns text.
The Price Tag That Changes the Math
Meta is leaning hard on one benchmark: Artificial Analysis's Intelligence Index, which scores models on reasoning quality and divides by cost. On that index, Ember's numbers look startling next to its bigger sibling.
| Model | Intelligence Index | Cost per task |
|---|---|---|
| Muse Ember | 36 | $0.14 |
| Muse Spark 1.3 | 37 | $2.39 |
| Claude Haiku 5.5 | ~36 (matched by Ember) | About 3× Ember's cost |
Read that again: Ember scores 36, one point behind Muse Spark 1.3's 37 — but costs $0.14 per task against Spark's $2.39. That is roughly 17 times cheaper for nearly identical benchmark performance. Meta also says Ember beats OpenAI's GPT-6 Luna at every reasoning setting, and that no model on the full Artificial Analysis leaderboard scores higher for less money per task.
The pricing behind those figures: $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. For developers running agents at scale — thousands of tool calls an hour — that gap is the difference between a feature that ships and one that never leaves the lab.
The Honest Caveat
Every number above comes from Meta's own announcement and from Artificial Analysis's composite index, so treat the framing as vendor-backed until more independent testing lands. And there is one result Meta is not shouting about: on real coding tasks, Claude Haiku 5.5 still wins. In the FrontierCode 2.0-AA coding benchmark, Haiku scored 34.9% against Ember's 27.2%.
That split tells you what Ember is for. If the job is writing hard code from scratch, the bigger models still earn their price. If the job is routing, extraction, summarising long documents or running the same small task ten thousand times, Ember does it for pennies — and that is most of what agents actually do.
What It Can Actually Do
Ember's headline specs are built for volume work. Its 1-million-token context window can swallow an entire codebase or document set in a single call, and it can write up to 512,000 tokens in reply. It handles text, images, audio and video as input, so an agent can feed it a screenshot, a voice note or a PDF and get structured text back.
Meta says it evaluated Ember under its Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks. It is available now through the Meta Model API and Muse Code, Meta's coding agent, under the name muse-ember-1.0, and it is also listed on OpenRouter and available on Google Cloud and Oracle Cloud.
Where It Sits in the Muse Family
| Model | Launched | Role |
|---|---|---|
| Muse Spark | April 2026 | Mid-weight workhorse |
| Muse Glimmer | August 2026 | Open-source twin of Spark |
| Muse Wonder | October 2026 | Frontier flagship |
| Muse Ember | November 2026 | Smallest, cheapest, built for agent work |
The pattern is now clear: Meta is building a ladder, not a single model. Wonder takes the hardest jobs, Spark handles everyday work, Glimmer serves the open-source community, and Ember does the cheap, repetitive work at scale. The bet is that the cheapest rung is where most of the real usage will land.
How to Try It
- Use the Meta Model API — Ember is listed as muse-ember-1.0, priced at $0.10 per million input tokens and $0.50 per million output tokens.
- Or open Muse Code, Meta's coding agent, where Ember is available now for agent and coding work.
- Or try it on OpenRouter, which lists the model with live pricing and provider details.
There is no free consumer chatbot version — Ember is a developer model, so you will need an API account or a coding-agent subscription to use it. But at 10 cents per million tokens, a first experiment costs less than a cup of coffee.
What This Means for the Rest of Us
Muse Ember is a sign of where the AI race is heading: the fight is no longer only about who builds the smartest model, but who builds the cheapest one that is good enough. If agents are going to do our repetitive work — sorting email, triaging tickets, summarising documents — they need to cost almost nothing per task, or nobody will let them run.
Meta is betting that a tiny model trained by a giant one is the answer. The early benchmarks say the bet is credible. The real test comes now, as developers put Ember to work and find out whether 10 cents buys enough intelligence for the jobs that matter.
FAQ
Can I try Muse Ember for free?
There is no free consumer version. Ember is a developer model available through the Meta Model API and Muse Code, and listed on OpenRouter. At $0.10 per million input tokens, though, trying it costs almost nothing.
Is Muse Ember better than Claude Haiku 5.5?
On Artificial Analysis's Intelligence Index, Ember matches Haiku 5.5's best score at about a third of the cost per task. But Haiku still wins on real coding tasks — 34.9% vs 27.2% on FrontierCode 2.0-AA. For routine agent work, Ember is the better value; for hard coding, Haiku still leads.
Is Muse Ember open source?
No. The open model in Meta's Muse family is Glimmer, the open-source twin of Spark. Ember is a proprietary model you use through Meta's API, Muse Code or OpenRouter.
Sources: Meta — Introducing Our Smallest, Most Efficient Model: Muse Ember ↗ · TechCrunch — Meta releases Muse Ember, a new lightweight model for agent execution ↗ · The Decoder — Meta launches Muse Ember, a small AI model built for the repetitive work agents actually do ↗ · OpenRouter — Muse Ember API pricing & providers ↗
