---
title: "Muse Ember launches: what the benchmarks show for low-cost agent work"
description: "Compare Muse Ember, Claude Haiku 5.5 and GPT-6 Luna on three launch benchmarks, plus Ember’s availability, API pricing and long-prompt threshold."
url: "https://promptszone.com/muse-ember-launch-benchmark-comparison"
published: "2026-11-10T01:00:05Z"
categories: ["New Models"]
---

# Muse Ember launches: what the benchmarks show for low-cost agent work

Meta’s Muse Ember launches for agent workflows. Compare three benchmarks, launch API pricing and the long-prompt threshold before choosing a cheaper model.

![November 9 benchmark comparison of Muse Ember, Claude Haiku 5.5 and GPT-6 Luna at maximum reasoning, with row-leading values bolded.](https://promptszone.com/assets/uploads/670d27bc5074e88f5f0301acdbe89fc5.png)

**Summary:**

Meta’s smallest model is available through the Meta Model API and Muse Code. A three-benchmark comparison shows where Ember matches or trails peers, with practical checks for cost-sensitive agent workflows.

Meta released Muse Ember on November 9, positioning its smallest model for the repeated steps that make agent workflows expensive: routing requests, extracting information, calling tools and handling coding subtasks. The launch is an availability announcement, not just a research preview: Meta says Ember is live through the Meta Model API and Muse Code as **muse-ember-1.0**.

The interesting question is where a cheaper model can take over work without creating more retries or review. Ember’s launch benchmarks suggest a competitive general-purpose baseline, with meaningful differences between individual tasks.

## Three benchmarks, three different comparisons

The following November 9 launch snapshot uses Artificial Analysis results reproduced in Meta’s announcement. All three models use their maximum reasoning setting. Higher is better in every row; bold marks the best displayed value, including ties. The shaded Ember column in the cover chart identifies the subject, not a winner.

| Evaluation | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
| --- | --- | --- | --- |
| AA Intelligence Index v5 | **36** | **36** | 28 |
| FrontierCode 2.0 | 27.2% | **34.9%** | 25.2% |
| GDP.pdf | **22.8%** | 20.8% | **22.8%** |

The index values are rounded: Ember and Haiku share the displayed score of 36, rather than an identical underlying result. The rows measure different tasks and should not be averaged together.

Haiku leads this coding comparison, while Ember ties Luna on GDP.pdf and finishes above Haiku there. That makes workflow-level testing more useful than selecting a model from the overall index alone. Artificial Analysis’s separate launch review also identifies knowledge accuracy as a weaker area for Ember, so document extraction and open-ended factual answering should not be treated as interchangeable workloads.

## Low token prices, with a long-prompt threshold

Meta lists a 1-million-token context window and up to 512,000 output tokens. For prompts of up to 200,000 tokens, launch API prices are US$0.10 per million input tokens, US$0.004 per million cached input tokens and US$0.50 per million output tokens.

Above that prompt threshold, the listed rates rise to US$0.30, US$0.012 and US$1.50 respectively. A large context window therefore does not mean every request uses the lowest advertised rate.

Artificial Analysis reports roughly US$0.14 per benchmark task for Ember at maximum reasoning, versus US$0.41 for Haiku 5.5 at the same setting. Those are costs for its evaluation workload, not a forecast for an application with different prompts, output lengths or retry behavior.

## Where to start evaluating Ember

- **Start with a bounded repeated task.** Use a known extraction format, routing decision or coding subtask with an answer you can check.
- **Count successful completions, not just tokens.** Include invalid tool calls, retries and human corrections in the comparison.
- **Test long documents separately.** Track whether requests cross the 200,000-token pricing threshold and whether cached inputs actually qualify.
- **Keep difficult cases in the test set.** Include missing information, ambiguous instructions and questions that should produce an abstention rather than a guess.

Ember is worth testing as a lower-cost component in an agent system. The published results support that narrower proposition more clearly than a claim that it replaces a larger model across every task.

## Sources

- [Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember](https://research.meta.ai/blog/muse-ember), November 9, 2026. Launch availability, specifications, pricing and benchmark comparison.
- [Artificial Analysis: Meta’s Muse Ember launch analysis](https://artificialanalysis.ai/articles/meta-s-muse-ember-is-the-cheapest-model-scoring-36-on-the-intelligence-index), November 9, 2026. Independent evaluation and benchmark-task costs.
- [Artificial Analysis: Muse Ember model profile](https://artificialanalysis.ai/models/muse-ember). Live results can change after this dated launch snapshot.

Original benchmark chart by PromptsZone, using Artificial Analysis results reproduced in Meta’s launch announcement.