Meta Muse Ember Guide: Price, Benchmarks vs Haiku 5.5 and GPT-6 Luna, and How to Use the API (2026)

Meta has launched Muse Ember, the smallest and cheapest model in its Muse family. It went live on 9 November 2026 on the Meta Model API and in Muse Code.

Ember is built for the boring, high-volume half of agent work: tool calls, routing, data extraction, coding subtasks and long document parsing. Meta’s pitch is simple: “Muse Wonder plans, Muse Ember handles the volume.”

This guide covers the price, every benchmark Meta published, how it stacks up against Claude Haiku 5.5 and GPT-6 Luna, and how to call it from Python in five minutes.

What Muse Ember Is and Where It Fits

Meta Superintelligence Labs calls Ember its “smallest, most efficient model”. It is designed for “complex reasoning execution”: jobs where the rules are already clear and the model must apply them thousands of times, fast and cheap.

Ember is the fourth Muse model, and the first one trained using Muse Wonder Full. Here is how the family lines up today.

Meta lists the jobs Ember is made for:

• Agent workflows: pulling records, calling tools in sequence and updating systems across apps.
• Coding subtasks: reading files, running tests and making targeted fixes.
• Computer use: operating apps from screenshots.
• Long documents and media: charts, scans, PDFs, voice memos, meeting recordings and video, without a separate model.

Note: Ember is not the open model. If you want Muse weights on your own GPU, that is Muse Glimmer or Muse Spark Open. See my guide on running Muse Spark Open locally.

Specs and Pricing

All figures below come from Meta’s launch post and the Muse Ember page in the Model API docs.

Price per million tokens

Two extra costs to know. Reasoning tokens count as output, so a max effort call bills more than its visible answer. Search grounding on the Model API costs $2.50 per 1,000 queries on top of tokens.

Tip: Ember’s cached input is 25 times cheaper than fresh input. Put your long system prompt and tool definitions first and keep them identical, so repeat calls hit the cache. My prompt caching guide explains the pattern.

What a real job costs (estimate)

Say you classify 10,000 support tickets, each with 2,000 input tokens and 300 output tokens (including reasoning). That is 20M input and 3M output tokens.

• Input: 20M x $0.10 = $2.00
• Output: 3M x $0.50 = $1.50
• Total: about $3.50, before any cache savings

This is my arithmetic from list prices, not a measured run. Max-effort reasoning can push output tokens well above 300 per ticket, so test a sample first.

Benchmarks: Ember vs Haiku 5.5 and GPT-6 Luna

Meta compared Ember with the two other budget leaders, Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna. All three ran at max reasoning. Rows marked “Meta internal” are Meta’s own tests; “-AA” rows were run by Artificial Analysis. All scores are vendor-reported.

The big gap is on AutomationBench, which tests multi-step business workflows. That is the exact job Meta built Ember for. On pure hard coding (FrontierCode 2.0), Haiku 5.5 still wins clearly.

Ember also nearly matches its bigger sibling. On the Intelligence Index it scores 36 vs 37 for Muse Spark 1.3, while costing about $0.14 per task vs $2.39.

Warning: Three of these rows are Meta’s own internal benchmarks, and Meta’s Ember cost figures were calculated at “proposed pricing”. Wait for more independent runs before you swap a production model on these numbers alone.

Same Sticker Price, Lower Bill: Why Cost per Task Matters

Here is the twist most launch coverage skipped. Ember, Haiku 5.5 and GPT-6 Luna all list at $0.10 input / $0.50 output per million tokens for normal-size prompts.

So why does Meta call Ember “significantly cheaper”? Because the bill depends on how many tokens a model burns to finish a task, and where its long-prompt surcharge starts.

Artificial Analysis summed it up in a chart: Ember matches Haiku 5.5 at max reasoning for about a third of the cost per task. Early tester You.com also reported 36 to 48% fewer output tokens than leading open models on coding tasks.

TechCrunch adds that Ember is cost-competitive with the cheap open models served by API, DeepSeek V4.1 Flash and GLM 5.3 Flash. For current DeepSeek rates see my DeepSeek API pricing guide.

How to Use Muse Ember (API and Muse Code)

The Meta Model API speaks the OpenAI format, so you can use the official openai SDK. If you already have OpenAI code, you only change the base URL, the key and the model name.

Call it from Python

1. Create an account. Sign in at dev.meta.ai with a Meta Model API account and add billing (Ember is pay-as-you-go).
2. Make a key. In the dashboard open API keys and click Create API key.
3. Save it as an environment variable. On macOS or Linux add export MODEL_API_KEY="your-key" to ~/.bashrc or ~/.zshrc. On Windows add it under System, Environment Variables.
4. Install the SDK. Run pip install openai (or npm install openai for TypeScript).
5. Run the code below. Pass the key explicitly, because the OpenAI SDK does not read MODEL_API_KEY on its own.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

response = client.responses.create(
    model="muse-ember-1.0",
    reasoning={"effort": "low"},
    input="Classify this ticket as billing, bug or how-to: 'I was charged twice this month.'",
)
print(response.output_text)

The response holds a reasoning item and a message item. The SDK’s output_text gives you just the answer.

Tip: The default effort is max, which is slow and token-hungry. For routing and classification, try minimal or low first. Meta’s own photo-sorting sample calls Ember at minimal effort.

Quick test with curl

curl https://api.meta.ai/v1/responses -H "Authorization: Bearer $MODEL_API_KEY" -H "Content-Type: application/json" -d '{"model": "muse-ember-1.0", "input": "Summarise the Muse model family in one line."}'

Reliable extraction prompts

Meta’s cookbook uses a pattern worth copying for invoices, contracts and forms. Ask for each field with three parts:

• value: the extracted text, or null if it is not there
• source: the page label where it was found
• status: present, missing or unreadable

Combine it with structured output so the reply is always valid JSON. My structured outputs guide shows the schema approach.

Use it in Muse Code

Muse Code is Meta’s coding agent CLI. Ember is available on pay-as-you-go and every subscription plan. Install it from dev.meta.ai, then pick the model with the model flag or set Ember as the worker model in an agent definition.

Useful options from Meta’s samples include the model, reasoning-effort, agents and parallel-tool-calls flags. Be careful with the yolo option: it approves shell commands without asking, and agents can read your API key.

Note: On a Muse Code subscription, requests over 200K input tokens use up your plan allowance faster.

The Wonder + Ember Pattern for Agents

Meta’s recommended setup splits work between a planner and many cheap workers. Your app or harness decides which model handles each call.

1. Plan with Wonder or Spark. The big model writes the plan, the dependencies and the acceptance criteria. Each handoff lists the task, context, files the worker owns, checks and expected result.
2. Execute with Ember. Ember workers run the bounded subtasks in parallel and report blockers back.
3. Review with Wonder or Spark. The planner reasons through anything unresolved, alongside a human review.

Meta published two worked examples with real bills:

In the photo job, Ember’s albums agreed with the dataset labels about 86% of the time. That is decent for bulk sorting, and a reminder to keep a human check on anything important.

Ember vs Haiku 5.5 vs Luna vs Spark: Which to Pick

For a full breakdown of the cheap tier across vendors, see my Claude Haiku 5.5 guide and the AI API pricing comparison.

Limits, Privacy and What Meta Has Not Said

• No open weights. Ember is API-only. Meta does not say if a Glimmer-style open twin will follow.
• Bounded work only. Meta itself says Ember is for repeatable tasks and points hard problems to Wonder. Results also vary from run to run, per Meta’s build guide.
• Rate limits not published yet. The pricing page lists 3,000 requests per minute for Spark and Wonder, but no Ember figures at launch.
• No batch price. Meta lists none for Ember, unlike Anthropic’s 50% batch discount on Haiku.
• Regions. Meta’s docs do not list supported countries for the Model API.
• Privacy: Ember runs in Standard mode only. In Standard mode, Meta says your content is not used to train its models. The discounted Contributor mode, which does allow training, is not offered for Ember.
• Safety: Meta says Ember stayed below the high-risk thresholds of its Superintelligence Scaling Framework for chem-bio, cyber and loss-of-control. It was also tested on prompt injection and asking before high-impact actions.
• Meta AI app: Meta has not said whether Ember powers anything in the consumer Meta AI app.

Troubleshooting

My Take

Ember is the most interesting launch in the budget tier this autumn, but not for the reason in the headline. The price is not lower than Haiku 5.5 or GPT-6 Luna. What is different is how far that price stretches.

A 200K base-price window, a $0.004 cache read and audio and video input make Ember a strong fit for document-heavy agents and support pipelines. If your workflows look like AutomationBench, this is worth a trial this week.

I would not move a coding agent to Ember yet. Haiku 5.5 leads on hard code, and several of Meta’s best numbers come from internal tests. Run your own 100-sample eval, compare cost per finished task, and let that decide.

FAQ

How much does Muse Ember cost?

$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens, with cached input at $0.004. Above 200K tokens it is $0.30 / $1.50.

Is Muse Ember open source?

No. Ember is only available through the Meta Model API, Muse Code and partner clouds. Meta’s open-weight models are Muse Glimmer and Muse Spark Open.

What is the Muse Ember model ID?

muse-ember-1.0 on the Model API, at base URL https://api.meta.ai/v1.

Is Muse Ember better than Claude Haiku 5.5?

On Meta’s tests it beats Haiku on agent and workflow benchmarks and ties it on the Artificial Analysis index at about a third of the cost per task. Haiku 5.5 is still stronger on hard coding.

Can I use Muse Ember with the OpenAI SDK?

Yes. Point the SDK at https://api.meta.ai/v1, pass your Model API key and set the model to muse-ember-1.0.

Does Muse Ember accept audio and video?

Yes. It takes text, images, audio, video and PDF as input and returns text.

Will Meta train on my Ember prompts?

Ember only runs in Standard mode, and Meta says Standard-mode content is not used to train its models.

Sources

• Meta AI Research: “Introducing Our Smallest, Most Efficient Model: Muse Ember” (9 Nov 2026), research.meta.ai/blog/muse-ember: specs, pricing, benchmark tables, partners, safety
• Meta Model API docs: Muse Ember page, Quickstart, Overview and Pricing and rate limits (dev.meta.ai/docs): model ID, endpoints, effort levels, base URL, errors, Spark/Wonder prices
• Meta developer blog: “Start building with Muse Ember” (dev.meta.ai/resources/blog/build-with-muse-ember): Wonder + Ember pattern, sample costs, extraction prompt
• TechCrunch: “Meta releases Muse Ember, a new lightweight model for agent execution” (9 Nov 2026): positioning, DeepSeek/GLM comparison
• TestingCatalog: “Meta launches Muse Ember for low-cost agentic tasks”: Artificial Analysis cost-per-task chart
• VKTR: GPT-6 Sol and Luna launch pricing (22 Sep 2026); Anthropic Claude Haiku 5.5 pricing (via my Haiku 5.5 guide)