Muse Ember: what it is, what it costs and when to swap Haiku 5.5 or GPT-6 Luna for Meta's small model

What you get from this article

• Muse Ember is Meta's smallest model, for repetitive agent work: tools, coding subtasks, routing and extraction. It is the first model trained using Muse Wonder Full.
• It costs US$ 0.10 input and US$ 0.50 output per million tokens, the same as Claude Haiku 5.5 and GPT-6 Luna. Cached input costs US$ 0.004, 2.5 times less.
• Above 200 thousand input tokens, the whole request costs 3 times as much. On Haiku 5.5 the step starts at 100 thousand and multiplies by 5.
• In Meta's table, Ember wins 7 of 10 rows, ties 2 and loses FrontierCode 2.0 to Haiku 5.5 (34.9 against 27.2).
• In Meta's chart built on Artificial Analysis data, it gets close to Haiku 5.5's best score while spending about a third per task. The blog itself calls that cost both proposed and list pricing.
• On the API, Meta does not train on your content. On Muse Code subscriptions, the mode that trains on your content is on by default.
• Knowledge cutoff, batch pricing and a full model card had not been published by the night of the launch.

What Muse Ember is

Muse Ember is Meta's smallest language model, released on 9 November 2026 by Meta Superintelligence Labs. The company describes it as built for the work agents repeat all day: calling tools in sequence, running coding subtasks, routing requests, extracting data and reading long documents. The API identifier is muse-ember-1.0.

It accepts text, images, audio and video as input and answers in text. The context window is 1 million tokens, and a response can reach 512 thousand tokens, four times the output limit of Claude Haiku 5.5 and GPT-6 Luna. According to Meta, it is the first model trained using Muse Wonder Full, the larger version of the family released in October.

Meta arranges the family in three tiers: Muse Wonder for the hardest problems, Muse Spark for coding and agent workflows, and Ember for repeatable work at scale. The announcement spells out the pairing: Wonder plans and Ember handles the volume. It is the same agent design Anthropic suggested the same day with Fable 5.2: an expensive model that decides and cheap models that execute.

How much Muse Ember costs

On the Meta Model API, Ember costs US$ 0.10 per million input tokens, US$ 0.004 per million cached input tokens and US$ 0.50 per million output tokens, for requests of up to 200 thousand input tokens. That is exactly the base price of Claude Haiku 5.5 and GPT-6 Luna. The difference lies in caching, which costs 2.5 times less than on both rivals, and in the long-context steps.

Above 200 thousand input tokens, Meta bills the whole request at the higher rates, not only the tokens past the limit: US$ 0.30 input, US$ 0.012 cached and US$ 1.50 output, three times the base price. Cached tokens count toward the tier, and a conversation continued through the previous response identifier carries its history. On Haiku 5.5 the step comes earlier, at 100 thousand tokens, and multiplies the price by five. On GPT-6 Luna the jump comes above 272 thousand tokens and is smaller. Between 100 thousand and 200 thousand input tokens, Ember is five times cheaper than Haiku on input; between 200 thousand and 272 thousand, Luna is cheaper than Ember.

Rate limits are 3,000 requests and 4 million tokens per minute, per team. Web search, when used, costs US$ 2.50 per thousand queries on top of tokens. Meta's price table shows no batch price for Ember. In Muse Code, Meta's coding tool, subscriptions range from US$ 5 to US$ 200 per month.

Price in US$ per million tokens on each company's API

	Input	Cache	Output
Muse Ember up to 200 thousand tokens	0.10	0.004	0.50
Muse Ember above 200 thousand	0.30	0.012	1.50
Haiku 5.5 up to 100 thousand tokens	0.10	0.01	0.50
Haiku 5.5 above 100 thousand	0.50	0.05	2.50
GPT-6 Luna up to 272 thousand tokens	0.10	0.01	0.50
GPT-6 Luna above 272 thousand	0.20	0.02	0.75
Muse Spark 1.3 Meta, one tier up	1.25	0.15	4.25

Same price per token, different cost per task

With per-token prices tied, Meta makes its case for Ember on the cost of each completed task. The blog includes a chart with the Artificial Analysis intelligence index on the vertical axis and average cost per task on the horizontal one, one point per effort level. At max effort, Ember scored 35.8 at US$ 0.136 per task. Haiku 5.5 at max scored 36.2 at US$ 0.411, three times as much. GPT-6 Luna's best point was 27.8, at US$ 0.152.

Meta's conclusion is that Ember delivers more intelligence per dollar than both rivals at any effort level. On the published points, the claim holds against Luna. Against Haiku, Ember gets close to the top score, without passing it, while spending a third. For anyone running thousands of calls a day, that is the difference that shows up on the bill, because the list price is the same.

The blog itself carries a caveat. The chart caption says Ember's costs are estimates at proposed pricing, and the note just below says they were computed at list pricing. As of this writing, Meta had not explained the difference. And the index belongs to Artificial Analysis, which tests each model its own way; with your task and your prompt, the math may come out differently.

Artificial Analysis index and cost per task, from Meta's chart

	Index	Cost per task
Muse Ember xhigh Meta	33.9	US$ 0.116
Muse Ember max Meta	35.8	US$ 0.136
Haiku 5.5 high Anthropic	29.6	US$ 0.159
Haiku 5.5 xhigh Anthropic	34.3	US$ 0.248
Haiku 5.5 max Anthropic	36.2	US$ 0.411
GPT-6 Luna max OpenAI	27.8	US$ 0.152

Benchmarks: where Muse Ember wins and where it loses

The launch table compares Ember with Claude Haiku 5.5 and GPT-6 Luna, all at max effort. Across 10 rows, Ember leads outright on 7, ties on 2 and loses 1. The widest lead is in business workflows: 37.8% on AutomationBench, a private set of 657 Zapier tasks across apps such as Gmail, Salesforce and Slack, against 20.2% for Luna and 13.9% for Haiku.

The loss is on a test that matters to developers. On FrontierCode 2.0, which checks whether a code change would be accepted by the project maintainer, Haiku 5.5 scored 34.9 against Ember's 27.2. On computer use, the lead is 0.6 points (69.5 against 68.9), and on the overall Artificial Analysis index the two tie at 36.

Two rows need careful reading. On Terminal-Bench 4.0, each model was measured in a different setup: Ember by Artificial Analysis in a simple agent, Haiku by the number Anthropic itself published in Claude Code, and Luna through Codex. And the Personal Superintelligence Index, Meta's internal index, now combines eight evaluations; at the Muse Wonder launch in October it combined seven. The two scores are not comparable.

Muse Ember against Claude Haiku 5.5 and GPT-6 Luna

	Muse Ember Meta	Haiku 5.5 Anthropic	GPT-6 Luna OpenAI
AutomationBench-AA business workflows	37.8	13.9	20.2
Agents' Last Exam end-to-end agent tasks	61.6	57.0	57.1
Personal Superintelligence Index Meta internal index	46.8	36.1	37.0
Terminal-Bench 4.0 terminal coding	40.9	39.2	16.4
Coding Agent Index-AA coding inside an agent	46	37	31
FrontierCode 2.0-AA code accepted by maintainers	27.2	34.9	25.2
GDP.pdf-AA work over long documents, tie	22.8	20.8	22.8
OSWorld 2.1 computer use, partial score	69.5	68.9	50.5
PerceptionBench visual perception, with tools	80.6	78.6	62.8
Artificial Analysis Index overall index, tie	36	36	28

What Muse Ember is for

The use Meta describes is the subagent: the model a larger agent calls hundreds of times per session to read a file, run a test, make a targeted fix or pull a field out of a document. Since it accepts images, audio and video, it can also read screenshots, charts, scanned documents, meeting recordings and voice memos without going through another model first.

Partners that tested it before launch reported these results, measured by themselves:

In a company's daily work, the natural candidates are message triage and routing, data extraction from invoices, contracts and forms, customer service summaries, lead classification and the subtasks of a coding agent. These are high-volume, short-answer jobs, where the price per call weighs more than the last decimal point of quality.

• Box: on the company's internal complex work evaluation, Ember scored a contract against the negotiation playbook, clause by clause, and reached the correct overall risk rating.
• Cloudflare: it ran up to 30% faster at the tail than the comparable model tested, on investigations across several repositories.
• You.com: it completed every coding task in the test with a p95 of 2.2 seconds, 4 to 13 times faster than leading open-weight models, and with 36% to 48% fewer output tokens.

Where to use it and what to check in the contract

Ember has been on the Meta Model API and in Muse Code since launch, and Meta said the same day that Ember, Spark and Wonder are now on Google Cloud and Oracle Cloud. The API supports the Responses format, and Meta's quick start tells developers to point an OpenAI SDK compatible client at its endpoint, which cuts the work for anyone with an existing integration. Meta has not published open weights for Ember; in the family, the open releases are Muse Glimmer and Muse Spark Open.

The point to watch is how your data is used. Meta has a mode called Contributor, which gives a discount in exchange for permission to train future models on what you send: prompts, code, files and responses. On the API, Ember runs only in Standard mode, where content is not used for training. On Muse Code subscriptions, the documentation says Contributor mode is on by default and can be turned off in settings.

For a Brazilian company sending customer data, that goes into the LGPD analysis, Brazil's data protection law, before it goes into the cost one. Through the API, the path is easier to document. On Muse Code, check the setting on every account in the team.

Pros of Muse Ember

The measured gains are in volume and long documents. The figures below come from Meta or its partners.

The easiest gain to measure is caching: in an agent that resends the same context at every step, it weighs more than the input price.

• Caching 2.5 times cheaper: US$ 0.004 per million, against US$ 0.01 for Haiku 5.5 and GPT-6 Luna.
• Cheaper mid-length context: between 100 thousand and 200 thousand input tokens, it costs a fifth of Haiku 5.5 on input.
• Long responses: up to 512 thousand output tokens, four times the limit of both rivals.
• Business workflows: 37.8% on AutomationBench against 20.2% for Luna and 13.9% for Haiku, in Meta's table.
• Multimodal input: text, images, audio and video in the same model.
• No training on API data: Ember runs only in Standard mode on the Meta Model API.

Cons and limits of Muse Ember

The limits appear in Meta's own documents and weigh more on teams that need predictability.

Information is also missing. As of the night of the launch, Meta had not published a model card for Ember, as it did for Wonder, nor its knowledge cutoff, available effort levels, batch price or the numbers from its safety evaluations, only that the model stayed below the high-risk thresholds of its framework.

• Production-quality code: it loses FrontierCode 2.0 to Haiku 5.5 (27.2 against 34.9), the test closest to code that ships.
• Long-context step: above 200 thousand tokens the whole request costs three times as much, and between 200 thousand and 272 thousand GPT-6 Luna is cheaper.
• Uneven comparison: on Terminal-Bench, each model was measured in a different setup.
• Cost per task with a caveat: the blog itself calls the price used in the chart both proposed and list pricing.
• Muse Code trains by default: Contributor mode is on in subscriptions.
• No per-language results: Meta has not published an evaluation by language.

Muse Ember, Haiku 5.5 or GPT-6 Luna: which to choose

With the same base price, the choice depends on the shape of your task. If your agent resends a lot of cached context or reads documents between 100 thousand and 200 thousand tokens, Ember has a clear price advantage. If input often goes past 200 thousand tokens, GPT-6 Luna and Haiku 5.5 deserve a side-by-side calculation. If the job is writing code that goes straight to human review, Meta's own table favors Haiku.

We have used Haiku 5.5 in the customer chat on this site since October. A support conversation fits comfortably under 100 thousand tokens, so for that use all three charge the same per token, and the decision would come down to cost per conversation and answer quality in Portuguese, which Meta has not published for Ember.

The test we recommend is the same as in our Claude Fable 5.2 article: pick 20 real tasks, run them on the three models at default effort, and compare total cost, accuracy and time. The tokens for each call come back in the API response. Independent scores for each model are on our AI Benchmark.

Where ROO3 comes in

Ember enters a race in which the three small models charge the same per token. Whoever measures pays less: which model finishes your task with fewer calls, fewer tokens and less rework?

At ROO3, we build agents and automations for companies and choose the model for each step with cost and accuracy tests, including each vendor's data settings. Our AI consulting runs that test on your real tasks.

If you want to know whether to switch the model you use today, or to build an agent with an expensive model that plans and a cheap one that executes, talk to us.

Frequently asked questions

What is Muse Ember?

It is Meta's smallest and cheapest language model, released on 9 November 2026. It is built for repetitive agent work such as calling tools, running coding subtasks, routing requests and extracting data from documents. It accepts text, images, audio and video, with a 1 million token window.

How much does Muse Ember cost?

On the Meta Model API, US$ 0.10 per million input tokens, US$ 0.004 cached and US$ 0.50 output, for requests of up to 200 thousand tokens. Above that, the whole request costs US$ 0.30, US$ 0.012 and US$ 1.50.

Is Muse Ember better than Claude Haiku 5.5?

In Meta's table, Ember beats Haiku 5.5 on 7 of 10 rows, ties on the overall Artificial Analysis index and loses the FrontierCode 2.0 coding test, 34.9 to 27.2. The base price is the same. Meta says it gets close to Haiku's best score while spending about a third per task.

Where can I use Muse Ember?

On the Meta Model API and in Muse Code, under the identifier muse-ember-1.0, and through Google Cloud and Oracle Cloud. The API supports the Responses format and OpenAI SDK compatible clients.

Does Meta use my data to train Muse Ember?

On the Meta Model API, Ember runs only in Standard mode, where content is not used for training. On Muse Code subscriptions, Contributor mode, which allows training on your content, is on by default and can be turned off in settings.

Does Muse Ember have open weights?

Meta has not published open weights for Ember: it is available as a service, through the API, Muse Code and partner clouds. In the family, the open releases are Muse Glimmer and Muse Spark Open.

Sources

• Meta, Introducing Our Smallest, Most Efficient Model: Muse Ember (9 Nov 2026)
• Meta, Muse Ember Evaluation Methodology (report, Nov 2026)
• Meta, Muse Ember model page
• Meta Model API, models
• Meta Model API, pricing and rate limits
• Anthropic, Claude API pricing
• OpenAI, API pricing
• @AIatMeta on X, Muse Ember announcement (9 Nov 2026)

Rodrigo Fávaro

Founder of ROO3, a marketing and technology agency in São José do Rio Preto, Brazil. Builds AI products running in production (Tobia, gerar.app, Pense Mercado) and maintains the AI Benchmark, a public ranking of AI models. See ROO3 AI consulting.

X @rodmf LinkedIn rodrigofavaro