On November 9, 2026, Meta introduced Muse Ember, the company's smallest and most efficient model to date. In AI agents, Muse Ember takes over the recurring small tasks: calling tools, forwarding requests, extracting data from documents, and analyzing long texts.
Meta offers access via the Meta Model API and via Muse Code, and the Muse models also run on Google Cloud and Oracle Cloud. Meta makes no statements about availability in Germany or the EU or about the license.
What can Muse Ember do, and what does the model cost?
Muse Ember processes text, images, audio, and video. Per request, the model can read up to one million tokens, and the response may comprise at most 512,000 tokens.
Meta states the prices per million tokens in the Meta announcement, in US dollars. Outputs cost five times as much as inputs, while Meta charges significantly less for cached inputs that stay the same across multiple requests. For inputs over 200,000 tokens, prices rise to three times the amount. Meta does not state a price in euros.
The comparison service Artificial Analysis measures Muse Ember in Max mode at 36 points on the Intelligence Index, on par with Claude Haiku 5.5 from Anthropic. On average, a test task with Muse Ember costs about a third of what Haiku 5.5 needs for it. In addition, Muse Ember uses about a third of the output tokens.
Wissenstest: Genauigkeit der Antworten
Omniscience-Test von Artificial Analysis, Anteil richtiger Antworten
Claude Haiku 5.534,7 %
Muse Ember21,1 %
Muse Ember lässt rund 64 Prozent der Fragen unbeantwortet
Quelle: Artificial Analysis
On Terminal-Bench 4.0, the model reaches 40.9 percent, while Haiku 5.5 gets 32.8 percent. On knowledge questions, however, Muse Ember falls behind: Accuracy in the Omniscience test is 21.1 percent, while Haiku 5.5 reaches 34.7 percent. The main reason: Muse Ember leaves around 64 percent of the questions unanswered. Artificial Analysis puts the hallucination rate at 17.4 percent, about half the value of Haiku 5.5.
On FrontierCode 2.0, Muse Ember scores 27.2 percent, behind Haiku 5.5 at 34.9 percent. Some results in Meta's announcement were measured by the company itself, and cost figures in Meta's chart are partly estimates based on proposed prices.
What this means for you
If you work exclusively in chat interfaces, you don't need to change anything: Meta lists only the API and Muse Code as access options. Muse Ember becomes interesting as soon as you build automations that sort emails, extract invoice data, or summarize documents. For that, a trial run with ten to twenty real examples from your daily work is worthwhile, because an index value says little about your task.
For questions that require specialized knowledge and for demanding programming tasks, a larger model remains the safer choice; the figures for Omniscience and FrontierCode point that way. If you send customer data through the interface, clarify data processing agreements and server location in Meta's terms beforehand, because the announcement is silent on the EU.
Meta introduced Muse Ember on November 9, 2026, and Artificial Analysis published the measurements on the same day.
Sources
