---
title: Meta launches Muse Ember: a tiny agent model that matches Claude Haiku 5.5 at a third of the cost
url: https://hernandez.systems/en/articles/meta-launches-muse-ember-a-tiny-agent-model-that-matches-claude-haiku-5-5-at-a-third-of-the-cost
lang: en
type: NewsArticle
section: AI
published: "2026-11-12T14:34:32+00:00"
modified: "2026-11-13T15:00:42+00:00"
publisher: Osky Networks
image: https://oskydj.sfo3.cdn.digitaloceanspaces.com/calamardo/news_item/2026-11-12/portada-bdf4de59.png
alternate_es: https://hernandez.systems/es/articulos/meta-lanza-muse-ember-un-modelo-chiquito-para-agentes-que-empata-claude
---

# Meta launches Muse Ember: a tiny agent model that matches Claude Haiku 5.5 at a third of the cost

> Meta introduced Muse Ember on November 9, 2026: the smallest, most efficient model in its Muse family, built for the work an agent repeats all day long: calling tools, solving coding subtasks, routing requests, extracting data, and reading long documents. It costs $0.10 per million input tokens and $0.50 per million output tokens, has a one-million-token context window, and, according to Meta's figures, matches or beats Claude Haiku 5.5 and GPT-6 Luna on a good chunk of agent benchmarks for a fraction of the cost. For anyone building bots, automations, or pipelines with subagents, the number that matters is the price per task, not the size of the model.

Meta [introduced Muse Ember](https://research.meta.ai/blog/muse-ember) on November 9, 2026: the smallest, most efficient model in its Muse family, built for the work an agent repeats all day long: calling tools, solving coding subtasks, routing requests, extracting data, and reading long documents. It costs **$0.10 per million input tokens and $0.50 per million output tokens**, has a one-million-token context window, and, according to Meta's figures, matches or beats Claude Haiku 5.5 and GPT-6 Luna on a good chunk of agent benchmarks for a fraction of the cost. For anyone building bots, automations, or pipelines with subagents, the number that matters is the price per task, not the size of the model.

## What Meta announced

Ember is the little sibling of Muse Spark and Muse Wonder. Meta describes it as a model for "complex reasoning execution" in the small, structured steps that eat up most of an agent's tokens. Here are the availability details from the announcement:

- **Identifier:** `muse-ember-1.0`, available since November 9 in Muse Code and the Meta Model API.
- **Price:** $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens.
- **Context:** up to one million input tokens and up to 512,000 output tokens, enough, according to Meta, to fit an entire codebase or a full set of documents in a single call.
- **Compatibility:** supports the Responses API.
- **Clouds:** Muse Ember, Muse Spark, and Muse Wonder are also available on Google Cloud and Oracle Cloud.
- **Natively multimodal:** it understands text, images, audio, and video, so an agent can read screenshots, charts, scans, voice notes, or recordings without going through another model.

There's an interesting origin detail: Meta says Ember is the first model trained using Muse Wonder Full, its largest model. When it introduced Wonder, it promised to use it to develop cheaper specialized models; Ember is the first delivery on that promise. The official recommendation is to combine them: "Muse Wonder plans, Muse Ember handles the volume."

## The numbers, according to Meta

The centerpiece of the announcement is a chart of intelligence versus cost per task with data from Artificial Analysis. Ember at its maximum reasoning level scores 35.8 points on the Artificial Analysis index at a cost of $0.136 per task. Claude Haiku 5.5 at its maximum level scores 36.2, but at $0.411: Ember is practically tied for about a third of the cost, which is exactly what Meta is bragging about. GPT-6 Luna doesn't go past 27.8 points at any level.

![Chart of intelligence versus cost per task: Muse Ember with 35.8 points at $0.136, Claude Haiku 5.5 with 36.2 at $0.411, and GPT-6 Luna with 27.8 at $0.152 at their maximum level](https://oskydj.sfo3.cdn.digitaloceanspaces.com/calamardo/news_item/2026-11-12/grafica-73127580.png)

On the individual benchmarks, all with reasoning at maximum, Meta's table looks like this:

| Benchmark | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| AutomationBench-AA (business workflows) | **37.8%** | 13.9% | 20.2% |
| Agents' Last Exam (internal) | **61.6%** | 57.0% | 57.1% |
| Terminal-Bench 4.0 | **40.9%** | 39.2% | 16.4% |
| Coding Agent Index-AA | **46** | 37 | 31 |
| FrontierCode 2.0-AA | 27.2% | **34.9%** | 25.2% |
| GDP.pdf-AA (long documents) | **22.8%** | 20.8% | **22.8%** |
| OSWorld 2.1, partial credit (internal) | **69.5%** | 68.9% | 50.5% |
| PerceptionBench with tools (internal) | **80.6%** | 78.6% | 62.8% |
| Artificial Analysis Index | **36** | **36** | 28 |

The biggest jump is on AutomationBench-AA, which measures end-to-end business workflows: almost triple Haiku. The benchmark Ember loses is FrontierCode 2.0, which grades real engineering tasks by whether the change could be merged; there Haiku beats it by more than seven points. Translation: Ember is good at executing, not necessarily at delivering the PR your team would approve without comments.

Against its own sibling, the difference is brutal. On the Artificial Analysis index, Muse Spark 1.3 scores 37 points and Ember 36, but Spark costs $2.39 per task and Ember $0.14, about 17 times less. Spark charges $1.25 per million input tokens and $4.25 per million output tokens.

## What the partners say

Meta published quotes from companies that tested it before launch. The most concrete ones are about speed. Cloudflare says Ember ran "up to 30% faster at the tail" than the comparable model they tested. You.com reports a p95 latency of 2.2 seconds in its programming tests and between 36% and 48% fewer output tokens. Cline, the coding agent, pitches it as a model to run as a subagent inside a harness, which is exactly the use case Meta is selling.

On Hacker News, where the announcement passed 300 points, Simon Willison [pointed out](https://news.ycombinator.com/item?id=50429024) that the model has six effort levels (minimal, low, medium, high, xhigh, and max) and that, in his informal test of drawing a pelican in SVG, more reasoning didn't translate into a better result. It's an anecdote, not a benchmark, but it's a good reminder that the maximum level isn't always worth what it costs.

## What to take with a grain of salt

My read is that this looks very good for anyone who pays for tokens at volume, but with three asterisks:

1. **These are Meta's figures.** Several benchmarks (Agents' Last Exam, OSWorld 2.1, PerceptionBench, and the Personal Superintelligence Index) are internal evaluations. The ones tagged "AA" come from Artificial Analysis, but the selection and presentation were done by Meta.
2. **Ember's cost in the chart is an estimate.** The chart itself says in one footnote that Ember's costs are estimated "at the proposed price" and, in another, that they use list price. With the price now published, it should match, but the real cost per task depends on how many reasoning tokens it consumes in your case.
3. **I haven't tested it.** Everything in this article comes from Meta's official announcement and the coverage cited below. The announcement talks about APIs and clouds; it doesn't mention downloadable weights.

On safety, Meta says it evaluated it with its Meta Superintelligence Scaling Framework (chemical, biological, cybersecurity, and loss-of-control risks) and that it fell below the high-risk thresholds. It also says it checked whether it resists instructions hidden in web pages, documents, or tool outputs, the famous prompt injection. For a model that will be reading emails and PDFs from strangers all day, that point matters more than any table, and it's precisely the one that's hardest to verify from the outside.

## What changes in practice

The pattern Meta is pushing is already common: a big model plans and a small, cheap model executes the hundreds of repetitive steps. What's new is that the small model no longer seems so far from the big one on agent benchmarks. If you run WhatsApp bots, document extraction flows, ticket classification, or coding subagents, Ember is a candidate to try in the part of the pipeline that eats the most tokens. The sensible thing is to measure it with your own cases before migrating: the table says that on code that gets merged without touch-ups it still falls behind.

## Sources

- [Introducing Our Smallest, Most Efficient Model: Muse Ember (Meta, official announcement)](https://research.meta.ai/blog/muse-ember)
- [Muse Ember Evaluation Report (Meta)](https://research.meta.ai/static/muse-ember-evaluation-report)
- [Meta lanza Muse Ember: una IA ultrabarata para programar y automatizar tareas con agentes (Hipertextual)](https://hipertextual.com/inteligencia-artificial/muse-ember-ia-programar-meta/)
- [Hacker News discussion](https://news.ycombinator.com/item?id=50429024)
