# Muse Ember gives AI agents a cheaper way to divide the work

**Summary:** Meta's Muse Ember combines low API prices with competitive agent benchmarks. It is worth testing for repeatable subtasks while another model handles harder planning.

- Canonical: https://markhuang.ai/news/muse-ember-cheaper-agent-work
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-09
- Section: News
- Tags: Meta, Muse Ember, AI Agents, Subagents, AI Pricing, Model Benchmarks
- Source: [Meta AI Research](https://research.meta.ai/blog/muse-ember)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![A compact amber-and-copper mechanism lifts blank paper slips from an open folio with several precision arms](https://cdn.markhuang.ai/news/muse-ember-cheaper-agent-work/hero.webp)

*Conceptual illustration generated with GPT Image 2: a small worker handles repeated document tasks. This is a metaphor for Muse Ember's role, not a depiction of Meta hardware.*

Checking a contract against a company's negotiation rules means working through individual clauses before deciding on an overall risk rating. Box says Meta's new Muse Ember can do that in its early evaluation, according to a partner account in [Meta's November 9 launch announcement](https://research.meta.ai/blog/muse-ember). That is the work Meta wants this model to take on: detailed tasks that an AI agent repeats.

Muse Ember is available through Meta Model API and Muse Code as `muse-ember-1.0`. Meta lists prices of $0.10 per million input tokens and $0.50 per million output tokens. Tokens are the units used to meter the material the model processes and produces. At those rates, a developer has a reason to try delegating routine document and coding work to Ember while keeping a more capable model for difficult planning.

The division of work is what interests me. When a workflow already has a clear plan or rulebook, a cheaper model that can follow it could make repeated tasks much more affordable.

## Writing the plan and applying it are different jobs

Meta describes Ember's work as tool calls, extracting information, reading long documents, and coding subtasks such as inspecting files, running tests, and making targeted fixes. An agent uses a model to choose actions and work through a task with software tools. A subagent is a worker given part of that task.

In the contract example, the company's negotiation playbook supplies the standard. The model applies it clause by clause. That is a more specific assignment than asking a model to invent the company's negotiating position. Box's account comes from a launch partner; it does not provide an independent measure of reliability.

Meta proposes pairing Ember with Muse Wonder: Wonder plans, Ember handles the volume. [TechCrunch's coverage](https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/) explains the same arrangement through customer support, with a heavier model establishing rules and a lighter one applying them across cases. The model that makes the plan does not have to carry out every step. A separate worker can take the repetitive assignments.

Ember also has a 1 million-token context window, the amount of material it can consider in a request. Meta says it understands text, images, audio, and video, including scanned documents and meeting recordings. The document work Meta describes extends beyond neatly typed text. Those are Meta's advertised capabilities; the benchmark results below do not establish equal quality across all those input types.

## The low price comes with meaningful benchmark evidence

Artificial Analysis's Intelligence Index combines evaluations of knowledge work, coding, science, and general capabilities. Its [independent model page lists Ember at maximum reasoning effort](https://artificialanalysis.ai/models/muse-ember) with a score of 36 and an average cost of about $0.14 per index task. The low price has a measured workload behind it.

Meta's more precise launch chart puts Ember at 35.8 points for $0.136 per task, against Claude Haiku 5.5 at 36.2 for $0.411 at its highest effort setting. That is nearly the same aggregate score for roughly one-third of the benchmark cost. It is a stronger comparison than putting token prices side by side, because it accounts for the tokens each model consumed during the evaluation.

The advantage depends on the assignment. AutomationBench-AA tests whether agents can complete business workflows through software applications. In Meta's published comparison at maximum reasoning, Ember scores 37.8%, versus 20.2% for GPT-6 Luna and 13.9% for Haiku. Meta's [evaluation report](https://research.meta.ai/static/muse-ember-evaluation-report) explains the strict grading: a task passes only when every objective is met and no guardrail is violated.

On FrontierCode 2.0-AA, which evaluates engineering changes for mergeability, the order changes: Ember scores 27.2%, behind Haiku's 34.9%. The same announcement makes a case for trying Ember on business workflows while leaving Haiku ahead on this test of code changes.

These scores also have boundaries. [Artificial Analysis describes its index](https://artificialanalysis.ai/methodology/intelligence-benchmarking) as primarily text-based and English-language, with speech and some other capabilities evaluated separately. A broad index helps choose candidates. It cannot settle how well Ember will understand a particular meeting recording or follow a company's contract rules.

## What a fifteen-cent token bill actually buys

For a hypothetical workload consuming 1 million uncached input tokens and 100,000 billed output tokens, Ember's published rates imply a $0.15 model bill: $0.10 for input plus $0.05 for output. That is arithmetic, not a measured contract-review cost. Tool charges, retries, and human review would sit outside it.

Meta also lists cached input at $0.004 per million tokens. Reusing eligible context could reduce the input bill further. Neither that discount nor Artificial Analysis's $0.14 average tells us the price of finishing a particular business job. I made the same distinction in [an earlier look at token pricing](https://markhuang.ai/news/token-sticker-price-is-a-trap): the useful comparison is the cost of work that passes review.

I would start Ember with a repeatable assignment whose results are easy to check, such as extracting specified fields from documents or making a narrowly scoped code fix. Compare it with the current model using the same inputs and tools, and count review time alongside the API bill. If Ember lowers that total at the required quality, move that class of work.

Muse Ember makes that experiment appealing. A team can try changing the model doing repeated subtasks while leaving difficult planning with the model already handling it. The contract example gives that division a concrete purpose: apply a known rulebook across the material, then bring the result back for review.
