Meta Muse Ember: Pricing and What It Means for Your Site

On November 9, 2026, Meta released Muse Ember, its smallest and cheapest AI model, built for repetitive agent work: classifying, extracting data, calling tools and reading long documents. It costs $0.10 per million input tokens. It matters to any business that runs, or plans to run, AI on its website.

What we know

What it is: Meta's research blog calls it its "smallest and most efficient model." Meta pitches it for bounded, repeated work (tool calls, coding subtasks, routing, extraction and long-document parsing) and suggests pairing it with Muse Wonder for the hardest problems.
Pricing: the official model page lists $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens for prompts up to 200,000 tokens. Above that, it is $0.30, $0.012 and $1.50.
Specs: model ID `muse-ember-1.0`, a 1-million-token context window, and text, image, audio and video input, according to Meta. Its developer guide says it works with an OpenAI-compatible client.
Where it runs: since November 9, on the Meta Model API and in Muse Code, according to the announcement. The model page also lists OpenRouter and Google Cloud. OpenRouter's listing says it is not currently available through EU in-region routing.
Meta's performance claims: Meta says it delivers more intelligence per dollar than Claude Haiku 5.5 and GPT-6 Luna, and matches Haiku's best Artificial Analysis score at about a third of the cost. Several of the benchmarks Meta publishes are internal or partner-reported.
The independent measurement: Artificial Analysis scores it 36 on its Intelligence Index, the same as Claude Haiku 5.5, at $0.14 per task versus $0.41. It also flags the weak spot: on its knowledge test, Muse Ember is right 21.1% of the time (Haiku: 34.7%) and declines to answer 64% of the questions.

What changes and what doesn't

Small tasks get cheaper. For a job with 20,000 input tokens and 3,000 output tokens (say, reading a multi-page PDF and returning its data as a table), this article's estimate using the official prices is about $0.0035 with Muse Ember and $0.14 with Muse Wonder. At 1,000 tasks a month, that is roughly $3.50 versus $140. It is an estimate: reasoning is billed as output and varies by task.
A small model still doesn't know much on its own. The independent test shows weak recall of world facts. It is built to work with the material you hand it, not to answer general-knowledge questions or facts about your business it has never seen.
No word on open weights. The announcement only mentions hosted access through Meta's API, Muse Code and clouds such as Google Cloud and Oracle Cloud; it offers no download.

How to tell if it affects you

List the repetitive jobs on your site. Sorting contact-form messages, pulling data from invoices or product sheets, summarizing reviews or routing inquiries to the right team are the kind of work Meta built it for.
Run the numbers before switching. Multiply the tokens a typical task uses by the prices above and by your monthly volume. Watch prompts over 200,000 tokens: the price triples.
Put your content in the prompt. If it powers an assistant on your site, include the page, product sheet or FAQ it should rely on in every request. Don't expect it to know your prices, hours or services.
Test with real cases. Take 20 or 30 examples your current model handles today and compare answers one by one, including format (for example, that JSON output is always valid), before moving all traffic.
Check data handling and region. The launch pages don't say where requests are processed or detail how data is used. If your site handles customers' personal data, read the terms of the Meta Model API or the provider you use before sending anything.

Sources

Updates: if Meta changes pricing, discloses where data is processed or brings the model to more platforms, it will be added here with a link.