Skip to content
Neural Daily

ModelsNews

Muse Ember: Meta's 10-cent model for agent work

🇭🇺 Olvasd el magyarul →

Meta released Muse Ember, the smallest and cheapest model in its Muse family. It costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1M-token context window. On the Artificial Analysis index it matches Claude Haiku 5.5 at about a third of the cost.

· 3 min read

A Muse Ember bejelentésének borítóképe: kék ívekből álló geometrikus minta világos háttéren
Image: Meta · Press kit · Source

What Meta released

Meta Superintelligence Labs introduced Muse Ember on November 9. Meta says it is built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing (Meta).

It is the fourth Muse model after Spark, the open Glimmer and the frontier Wonder. TechCrunch describes the intended setup in Meta's words: Muse Wonder plans, Muse Ember handles the volume (TechCrunch). Ember is the first model trained using Muse Wonder Full.

Pricing and specs

  • Price: $0.10 per million input tokens, $0.50 per million output tokens, $0.004 for cached input (three times higher for prompts over 200K tokens).
  • Context: 1M tokens, up to 512K output tokens.
  • Input: text, images, audio and video without a separate model.
  • Access: Meta Model API and Muse Code as muse-ember-1.0, with Responses API support. Also on Google Cloud and Oracle Cloud.

What the numbers show

On the Artificial Analysis Intelligence Index v5 at max reasoning it scores 36, the same as Claude Haiku 5.5, while GPT-6 Luna scores 28. Meta puts its cost at about $0.14 per AA task, versus $2.39 for Muse Spark 1.3, which scores only one point higher (37).

On agent tasks it scores 37.8% on AutomationBench-AA, against 13.9% for Haiku 5.5 and 20.2% for GPT-6 Luna. It reaches 40.9% on Terminal-Bench 4.0 (Haiku: 39.2%) and 69.5% on OSWorld 2.1 with partial credit. It does lose somewhere: on FrontierCode 2.0-AA, Haiku 5.5's 34.9% beats Ember's 27.2%, so it is not the best pick for complex, mergeable code changes.

Partners weigh in too: Cloudflare saw up to 30% better tail latency than a comparable model, and You.com measured a 2.2-second p95 on coding tasks. These are quotes Meta selected.

Why it matters

In agent systems most of the cost goes to many small steps: reading files, running tests, extracting data. That is where a big model planning and a cheap model executing pays off. Haiku 5.5 has led value rankings so far (AI value: Haiku 5.5 wins per dollar, Sonnet 5.5 at the top), and Ember targets that spot by Meta's numbers. Most figures come from Artificial Analysis, but PSI and OSWorld are internal, so test it on your own subagent tasks before swapping models.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles