---
type: "post"
id: "meta-launched-muse-ember-its-smallest-model-built-4baf24"
slug: "meta-launched-muse-ember-its-smallest-model-built-4baf24"
title: "Meta launched Muse Ember, its smallest model, built for agent subtasks and tool calls."
description: "Meta has released Muse Ember, a multimodal model for repeated agent work, on the Meta Model API and in Muse Code."
retrieval_nugget: "Meta has released Muse Ember, a multimodal model for repeated agent work, on the Meta Model API and in Muse Code. Its published token prices and task-level comparisons make it worth testing as a subagent for high-volume work, though benchmark results vary by task and do not establish performance in a particular production workflow."
published_at: "2026-11-09"
updated_at: "2026-11-09"
record_date: "2026-11-09"
date_kind: "published_at"
topics: ["claude"]
entities: ["Meta","Muse Ember","Meta Model API","Muse Code","Muse Wonder Full","Muse Spark","Claude Haiku 5.5","GPT-6 Luna"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "Meta introduced Muse Ember, its smallest, most efficient model for agent subtasks and tool calls, with cost-performance claims against Claude Haiku 5.5 and GPT-6 Luna."
prior_refs: []
source_url: "https://research.meta.ai/blog/muse-ember"
source_urls: ["https://research.meta.ai/blog/muse-ember"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:meta-launched-muse-ember-its-smallest-model-built-4baf24"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/meta-launched-muse-ember-its-smallest-model-built-4baf24/","markdown":"https://newruntime.com/posts/meta-launched-muse-ember-its-smallest-model-built-4baf24.md","json":"https://newruntime.com/posts/meta-launched-muse-ember-its-smallest-model-built-4baf24.json"}
markdown_url: "https://newruntime.com/posts/meta-launched-muse-ember-its-smallest-model-built-4baf24.md"
---

# Meta launched Muse Ember, its smallest model, built for agent subtasks and tool calls.

## Retrieval answer

Meta has released Muse Ember, a multimodal model for repeated agent work, on the Meta Model API and in Muse Code. Its published token prices and task-level comparisons make it worth testing as a subagent for high-volume work, though benchmark results vary by task and do not establish performance in a particular production workflow.

## Summary
Meta has released Muse Ember, a multimodal model for repeated agent work, on the Meta Model API and in Muse Code. Its published token prices and task-level comparisons make it worth testing as a subagent for high-volume work, though benchmark results vary by task and do not establish performance in a particular production workflow.

## What happened
Meta describes Muse Ember as its smallest and most efficient model, built for tasks agents repeat throughout a session: tool calls, coding subtasks, routing, extraction, and long-document parsing. The model is available as `muse-ember-1.0` on the Meta Model API and Muse Code, and supports the Responses API. Meta says it accepts text, images, audio, and video, and can handle a context window of up to 1 million tokens and output of up to 512,000 tokens.

The API price listed in Meta’s announcement is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Meta’s comparison with Muse Spark 1.3 lists Ember’s price as $0.10 input and $0.50 output per million tokens, against Spark’s $1.25 and $4.25. The announcement also cites an Artificial Analysis cost-per-task figure of about $0.14 for Ember, compared with $2.39 for Spark 1.3. That task estimate is distinct from token pricing and should not be treated as a quote for every workload.

Meta publishes results across agent, coding, multimodal, and general evaluations. At maximum reasoning, the company lists Muse Ember at 37.8% on AutomationBench-AA, compared with 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5. On Agents’ Last Exam, the figures are 61.6%, 57.1%, and 57.0%, respectively. Meta labels the former as strict pass@1 and says the latter uses its Linux subset.

Coding results are mixed. Meta lists Ember at 40.9% on Terminal-Bench 4.0, slightly above Haiku 5.5 at 39.2%, while GPT-6 Luna scores 16.4%. On FrontierCode 2.0-AA, however, Haiku leads with 34.9%; Ember scores 27.2% and Luna 25.2%. Meta says the benchmark uses real engineering tasks judged on mergeability. In multimodal tests, Ember and Luna both score 22.8% on GDP.pdf-AA, while Haiku scores 20.8%; Ember scores 69.5% on OSWorld 2.1 with partial-credit reward, against Haiku’s 68.9% and Luna’s 50.5%.

Meta says Ember is the first model trained using Muse Wonder Full and positions it for repeatable work at scale, while Muse Wonder is for the hardest problems and Muse Spark for coding and agentic workflows. The company also says its pre-release safety evaluations found Ember below its high-risk thresholds and that it received the same core safety training and testing as its other models.

## What the archive adds
New Runtime’s earlier coverage described Meta’s Muse Glimmer as a 30-billion-parameter open model for local agent tasks on consumer hardware ([October 20, 2026](https://newruntime.com/posts/meta-launches-muse-for-small-business-with-new-336903/)). This is a new model launch, not a new fact about Glimmer: Ember is presented as an API-available, low-cost model for repeated agent tasks. The linked earlier note concerns Muse for Small Business, a distinct Meta agent product, rather than Ember.

## Trend signal
This supports the pattern that agent economics is moving toward cost per completed task rather than cost per token or model call. Meta publishes both token prices and an estimated Artificial Analysis cost per task, alongside scores for agent-oriented tests. The distinction matters: a low token rate alone does not show what a successful task costs. [Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/).

## Why now
Meta says it is expanding a portfolio of specialized models developed after introducing Muse Wonder and Muse Wonder Full. Ember is the first model the company says was trained using Muse Wonder Full. Its stated rationale is that bringing capable models to more users requires a price that can scale across repeated product work.

## Why it matters
The useful question for agent builders is not whether a model wins a broad leaderboard, but whether it can complete a defined subtask at an acceptable cost, speed, and error rate. Ember’s prices and large context window create a candidate for routing high-volume work—such as extraction, document review, or routine coding steps—to a cheaper model while reserving a stronger model for planning or difficult cases.

Meta’s own results argue against switching on the basis of one headline score. Ember leads the listed comparisons on some agent and coding evaluations, but loses to Haiku on FrontierCode 2.0-AA, and ties Luna on GDP.pdf-AA. Those differences are a reminder to test against a team’s actual task mix and success criteria. A practical evaluation should include task completion, tool-call correctness, latency, retries, and the cost of failures, not just token price. Teams also need to check whether their API setup supports the model’s context and output limits and whether those limits matter for the workload they intend to route.

## Limits and uncertainty
The performance figures in Meta’s announcement are company-published comparisons, and some are identified as internal evaluations; they are not independent confirmation of results in production. The announcement does not establish how Ember performs on any particular team’s prompts, tools, or data. The $0.14 task figure is an estimate tied to Artificial Analysis, not a universal per-task price. Meta reports safety evaluations and that Ember remained below its framework’s high-risk thresholds, but those statements do not independently establish behavior across every deployment or tool environment.

## Skill to build now
Build a small routing evaluation before moving live traffic. Choose a representative set of routine tasks, record the expected result and acceptable error conditions, and run the current model and Ember through the same tools and inputs. Compare completion rate, review burden, latency, retries, and full API cost. Keep difficult or high-impact work on a path with human review until the lower-cost route has passed the team’s own checks.

## Reader takeaway
Muse Ember is available now with low listed token prices and a 1-million-token context window, aimed directly at repeated agent subtasks. Its benchmark profile is promising but uneven. Treat it as a candidate to test in your own harness, not as evidence that a model switch will automatically lower the cost of reliable completed work.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)