Skip to content
Main Site News Console

Introducing Our Smallest, Most Efficient Model: Muse Ember

· Simon Willison Translated
技巧LLM

Introducing our smallest, most efficient model: Muse Ember

Meta AI has just released Muse Ember, a model aimed at the same tier as Claude Haiku and GPT Luna. Here is how they describe it:

[…] our smallest, most efficient model, built for complex reasoning execution and the work agents repeat all day long: tool calling, coding subtasks, routing, information extraction, and long-document parsing.

With shorter prompts, it costs the same as GPT-6 Luna and Claude Haiku 5.5: $0.10 per million input tokens and $0.50 per million output tokens. Claude Fable 5.2, the other major model released today, costs 80 times as much, at $8/$40.

As prompts get longer, though, pricing becomes less straightforward. Luna, Ember, and Haiku all support up to one million input tokens, but each has a threshold above which the price goes up. Once that threshold is factored in, Ember’s pricing lands between the other two:

ModelPrompt sizeInput ($/million tokens)Cached inputOutput
GPT-6 Luna≤272k$0.10$0.01$0.50
GPT-6 Luna>272k$0.20$0.02$0.75
Muse Ember 1.0≤200k$0.10$0.004$0.50
Muse Ember 1.0>200k$0.30$0.012$1.50
Claude Haiku 5.5≤100k$0.10$0.01$0.50
Claude Haiku 5.5>100k$0.50$0.05$2.50

But simply comparing per-token prices doesn’t mean much anymore. With reasoning tokens, the number of tokens different models use to solve the same problem can vary enormously.

One useful yardstick here is Artificial Analysis.

A stacked bar chart from Artificial Analysis comparing the cost (in USD) of six models. Each bar is split into four unlabeled stacked segments, ordered from lowest to highest total cost. Claude Haiku 5.5 (low): total $0.04*, bottom segment $0.02, with three thinner unlabeled segments above. GPT-6 Luna (non-reasoning): total $0.07, bottom segment $0.06, with thinner unlabeled segments above. GPT-6 Luna (xhigh): total $0.10, with $0.07, $0.02, and $0.02 from bottom to top, plus a thinner unlabeled segment on top. Muse Ember (xhigh): total $0.12, with $0.06, $0.02, and $0.02 from bottom to top, plus a thinner unlabeled segment on top. Muse Ember (max): total $0.14, with $0.07, $0.03, $0.03, and $0.02 from bottom to top. Claude Haiku 5.5 (xhigh): total $0.25*, with $0.14, $0.04, $0.05, and $0.02 from bottom to top. Anthropic's bars are terracotta, OpenAI's are dark gray, and Meta's are blue, with each company's logo beneath its bars.

Even more useful is their “Intelligence Index vs. Cost per Task” chart, which shows these three models below:

A scatter plot from Artificial Analysis titled "Intelligence Index vs. Cost per Task". The vertical axis is the Artificial Analysis Intelligence Index (5 to 40), and the horizontal axis is cost per task in USD on a logarithmic scale ($0.04 to $0.3). A green "most attractive quadrant" sits in the upper left, along with a dashed Pareto line. Claude Haiku 5.5 (low)* scores about 19 at a cost of about $0.044. GPT-6 Luna (non-reasoning) scores about 9.5 at about $0.07. GPT-6 Luna (xhigh) scores about 23.5 at about $0.10. Muse Ember (xhigh) scores about 34 at about $0.12, and Muse Ember (max) scores about 36 at about $0.14; both fall on the Pareto line. Claude Haiku 5.5 (xhigh)* scores about 34.5 at about $0.25, well to the right of the Pareto line. Three additional unlabeled gray dots also sit on the Pareto line.

The Muse Ember models sit right on the Pareto line, meaning that along this line, improving one dimension requires sacrificing the other. Meta points this out specifically in its announcement:

It also sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model achieves a higher score at a lower cost per task.

(The gray dot in the middle of the green region, on the Pareto line, is GPT-6.1 Terra (low). Apparently I should be paying more attention to the Terra series!)

Muse Ember has six reasoning levels: minimal, low, medium, high, xhigh, and max, so I rendered six pelicans riding bicycles. The “high” reasoning level failed to return a complete SVG, and among the others, surprisingly, there was no clear correlation between reasoning level and the visual quality of the output.

Here is the result from the “max” level, though the others are worth a look too:

The bicycle frame is pretty good, there's a fish in the basket, and the pelican is wearing a strange blue hat.

The pelicans Ember generates are comparable in quality to the results I got from Muse Wonder, and Muse Wonder costs 40 times as much as Muse Ember.