How to Use Muse Ember: Meta’s Cheapest AI Model

Meta launched Muse Ember on 9 November 2026, and it is the cheapest model in the Muse family by a wide margin: $0.10 per million input tokens and $0.50 per million output tokens. It is the fourth Muse model, after Spark (April 2026), the open-weights Glimmer (August 2026) and Wonder (October 2026), and it is built for one job — the repetitive work AI agents do all day. Tool calls, coding subtasks, routing, extraction and long-document parsing. This guide shows when Ember is the right call, how to start using it, and where you should still reach for a bigger model.

Quick answer: Meta's Muse Ember costs $0.10 per million input tokens and matches Claude Haiku 5.5 on the Artificial Analysis index. A practical guide to using it in 2026.

What Muse Ember Is (and What It Is Not)

Muse Ember is a developer model, not a consumer app feature. It is available from 9 November 2026 on the Meta Model API and in Muse Code, Meta’s coding agent, under the model name muse-ember-1.0. It also appears on OpenRouter, and Meta says Ember, Spark and Wonder are now available on Google Cloud and Oracle Cloud as well. It supports the Responses API, takes text, images, audio and video, and accepts up to one million tokens in a single prompt, with outputs up to 512,000 tokens.

Two things it is not: it is not open-weights — Meta is keeping the weights private, as it did with Spark and Wonder (only Glimmer was released openly) — and it is not the model in the consumer Muse app. Meta has not announced Ember for the app, so do not expect it in your WhatsApp chats. It is also the first model Meta says was trained using Muse Wonder Full, its frontier model, which is how the company frames the small-model-versus-big-model trade-off: Wonder does the hard thinking, Ember inherits the distilled version for routine work.

Meta announced it in the plainest possible terms. As Mark Zuckerberg put it in his 9 November post on X:

“Releasing Muse Ember today — our smallest model that delivers higher intelligence per dollar than others in its class. Built for token-heavy work that agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.” — Mark Zuckerberg, X, 9 November 2026

Five companies tested Ember before launch — Box, Cline, Oracle, Cloudflare and You.com — which tells you the intended audience: teams running agents at volume, not individual chat users. If you are comparing the consumer app itself against its rivals, our Muse AI vs Gemini comparison covers that side of the family, and our guide to Muse Voice covers the app’s real-time voice feature — a separate product from this API model.

Muse Ember Pricing, Explained

The price list is short, and the interesting number is not the headline rate:

Token type Price per million tokens
Input $0.10
Cached input $0.004
Output $0.50

Source: Meta, “Introducing Our Smallest, Most Efficient Model: Muse Ember”, 9 November 2026.

The per-token rates are identical to Claude Haiku 5.5 and GPT-6 Luna on their standard tiers. Ember wins on what happens around the rate card. Its cached-input price is roughly 60% cheaper than Haiku’s, according to Artificial Analysis, and it generates about a third as many output tokens per task as Haiku 5.5 — around 84,000 against roughly 226,000 on the Artificial Analysis Intelligence Index. A request with 100,000 input tokens and 10,000 output tokens costs about $0.015; if every input token qualifies for the cached rate, the same request costs about $0.0054.

That is why the cost-per-task figures look so different from the price list. On the Artificial Analysis Intelligence Index v5, published 9 November 2026, Ember at its maximum reasoning setting costs about $0.14 per task. Haiku 5.5 costs about $0.41 for effectively the same score, and GPT-6 Luna about $0.15 for a much lower one. The bigger Muse models are in another league entirely: Muse Spark 1.3 costs $2.39 per task for one extra index point, and Muse Wonder costs about $6.95 per task. These are benchmark estimates from Artificial Analysis’s own prompts and measured token use, not production bills — your workload will vary — but the direction is clear: Ember is priced to be run everywhere, not rationed.

How Muse Ember Compares: The Benchmarks That Matter

Meta’s headline claim is that Ember “delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting”. The independent check comes from Artificial Analysis, which put Ember through its Intelligence Index v5 on launch day. Here is what the numbers actually say:

Benchmark Muse Ember Claude Haiku 5.5 GPT-6 Luna
Artificial Analysis Intelligence Index v5 (max reasoning) 36 36.2 28
Cost per Intelligence Index task ~$0.14 ~$0.41 ~$0.15
AutomationBench-AA (agent business workflows) 68.1% — —
GDP.pdf (document work) 22.8% 20.8% —
AA-Briefcase (Elo) 1,521 1,574 —
FrontierCode 2.0-AA (real coding tasks) 27.2% 34.9% —
AA-Omniscience 21.1% — —

Sources: Artificial Analysis, “Meta’s Muse Ember is the cheapest model scoring 36 on the Intelligence Index”, 9 November 2026; Meta’s launch evaluation tables, 9 November 2026; The Decoder, 9 November 2026. Dashes mean the figure was not reported in the sources used.

Read the table as a profile, not a victory lap. Ember ties Haiku on the headline index and beats it on agent workflows and document work, while costing a third as much per task — Artificial Analysis calls it the cheapest model scoring 36 or higher, and puts both its maximum and xhigh reasoning settings on the Intelligence-versus-Cost Pareto frontier. But Haiku is still ahead on real coding tasks (FrontierCode 2.0-AA, 34.9% against 27.2%) and on the Briefcase benchmark, and Ember’s 21.1% on AA-Omniscience is the lowest among models with a similar overall score. The honest summary: Ember is a strong cheap router and extractor, not a small Haiku. Use it where its strengths are, and the price does the rest.

How to Use Muse Ember in 5 Steps

Step 1 — Pick a repeatable job first. Ember is built for work your agents repeat hundreds or thousands of times a day. Start with one job you already run on a bigger model and measure the bill before and after. Do not start with your hardest reasoning task.

Step 2 — Get access. The two direct routes are the Meta Model API and Muse Code, both live from 9 November 2026 with the model name muse-ember-1.0. If your stack already routes through OpenRouter, Ember is listed there too. Meta’s announcement does not list the countries where the Meta Model API is available, so check availability for your region before you build on it.

Step 3 — Call it through the Responses API. Ember supports the Responses API, so a minimal call names the model and passes your input. Keep the shape simple at first:

{ "model": "muse-ember-1.0", "input": "Extract the invoice number, date and total from this document, and return them as JSON." }

Step 4 — Set the reasoning effort deliberately. Ember’s results are reported at two effort settings. Maximum reasoning scores 36 on the Intelligence Index at about $0.14 a task; the xhigh setting scores about 34 at about $0.12. That two-point gap is the cheapest quality dial in the Muse family — use maximum where accuracy decides the outcome, xhigh where volume decides the bill.

Step 5 — Cache what repeats, and set an escalation rule. Cached input costs $0.004 per million tokens, a fraction of the standard rate, so put your repeated system prompts and reference documents behind the cache. Then write down the rule for when Ember hands off: if it fails the same task twice, or the task needs long-form code or high-stakes judgement, escalate to Spark, Wonder or Haiku. A small model with a written escalation rule beats a big model used by default.

5 Jobs to Give Ember First

1. Tool-call routing and dispatch

The job Meta names first: deciding which tool an agent should call, and with what arguments. Short in, short out, repeated all day — exactly where Ember’s output-token efficiency pays.

2. Extraction from long documents

A one-million-token context window means a whole codebase or document set fits in one call. Pull invoice fields, contract clauses or log patterns out at volume, and keep the expensive model for reading what Ember flags.

3. Coding subtasks in Muse Code

Small edits, test generation and boilerplate inside Muse Code. Keep the FrontierCode gap in mind: Haiku still beats Ember on real coding tasks, so use Ember for the routine subtasks and escalate the hard ones.

4. Triage and classification at volume

Sorting support tickets, labelling content, routing messages. The output is a label or a short verdict, which is Ember’s cheapest shape of work.

5. Structured-output parsing

Turning messy agent outputs into clean JSON, summaries or tables. Short outputs, strict formats, easy to verify — and easy to catch when they go wrong. The same practical-first thinking runs through our guide to using Muse AI for your job search, where the workflow matters more than the model.

The 5-Question Ember Test

Before you route anything to Ember, run it through these five questions. If you answer yes to four or more, Ember is the right default. If you answer yes to two or fewer, keep the bigger model.

  1. Does this task repeat hundreds of times a day? Ember’s advantage is volume. A one-off hard question gains nothing from it.
  2. Is the output short? A tool call, a label, a snippet, a JSON object. Long-form writing is where bigger models earn their price.
  3. Is a wrong answer cheap to catch? Extraction you can validate, routing you can log, code you can test. If a mistake is expensive, the saving is false economy.
  4. Does the prompt repeat? Repeated system prompts and reference documents qualify for the $0.004 cached-input rate, which is where the bill really shrinks.
  5. Would a bigger model cost more than ten times as much for the same job? Against Spark 1.3, the answer is usually yes — $2.39 a task against $0.14.

Worked example: an agent that makes 10,000 tool calls a day, each with about 2,000 input tokens and 200 output tokens, uses 20 million input tokens and 2 million output tokens daily. On Ember that is $2.00 of input plus $1.00 of output — $3.00 a day, about $90 a month. The same workload on Muse Spark 1.3, at $1.25 per million input and $4.25 per million output tokens, costs $25.00 plus $8.50 — $33.50 a day, about $1,005 a month. Same calls, same volume, an eleven-fold difference. That gap is the whole point of Ember — and the reason to check, with the test above, whether your workload is actually the kind Ember is built for.

What We Don’t Know Yet

Ember is one day old, and the honest gaps matter as much as the launch numbers. Meta has not said which countries the Meta Model API serves, so availability outside the US is unverified. There is no independent replication of the benchmark results yet — the Artificial Analysis figures are independent measurements, but they were published on launch day, and the FrontierCode and Omniscience results come from Meta’s own evaluation report. Meta says it evaluated Ember under its Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks, but the full evaluation report is a separate document, not a summary in the announcement. And the obvious next question — whether Ember reaches the consumer Muse app — has no answer: Meta announced it for the API and Muse Code only. Treat all of this as the starting position, not the final verdict, and re-check before you build a product on it.

Our take: Ember is the most practical Muse launch of the year, precisely because it is boring. It does not try to be the smartest model — it tries to be the cheapest one that is good enough for the work agents actually repeat. If you run agents at volume, the routing question it forces — which jobs genuinely need a big model? — is worth more than the model itself. Most teams will find the answer is fewer jobs than they thought.

Frequently Asked Questions

Is Muse Ember free?

No. Ember is a paid API model. It costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens on the Meta Model API. There is no free tier announced.

Is Muse Ember in the Muse app?

Not that Meta has announced. Ember launched on the Meta Model API and in Muse Code, Meta’s coding agent. The consumer Muse app is a separate product, and Meta has said nothing about Ember coming to it. If you want to try Ember, you need API or Muse Code access, not the app.

How is Muse Ember different from Muse Spark?

Ember is smaller and far cheaper. It scores 36 on the Artificial Analysis Intelligence Index against Spark 1.3’s 37, but costs about $0.14 per benchmark task against Spark’s $2.39. Spark is the model for coding and agentic workflows where quality matters most; Ember is for repeatable coding and agentic tasks at scale.

Can I use Muse Ember on OpenRouter?

Yes. Ember is listed on OpenRouter as muse-ember-1.0, alongside the Meta Model API and Muse Code routes. Pricing through third-party routers may differ from Meta’s list prices, so check the router’s own listing.

What should I not use Muse Ember for?

Hard multi-step reasoning, long-form coding and high-stakes single-shot work. Haiku 5.5 still beats Ember on real coding tasks in FrontierCode 2.0-AA, and Ember’s AA-Omniscience score is weak. Use Ember for repeatable subtasks with cheap failure, and escalate the hard jobs to Spark, Wonder or Haiku.

Sources and Methodology

Claims in this article rest on Meta’s own launch announcement and Artificial Analysis’s independent benchmark article, both published 9 November 2026, cross-checked against launch coverage from The Decoder, MadRobot and Unite.AI. All prices and benchmark figures are as published on 9 November 2026; cost-per-task figures are Artificial Analysis estimates from its own prompts and measured token use, not production bills. The worked monthly-cost example is our own calculation from Meta’s published per-token prices. This article is desk-researched — Muse Ember was not tested hands-on, because API access and regional availability could not be verified first-hand at the time of writing. Verification date: 10 November 2026.

Leave a Reply

Your email address will not be published. Required fields are marked *