Meta launches Muse Ember, its cheapest small AI model for everyday agent work

What you could do with it

• A developer could build AI agents that make tool calls, route requests, and run coding subtasks all day at a fraction of the cost of earlier models ¹.
• A business could set up an automated system that reads and extracts information from long documents or PDFs without paying premium per-task prices ².
• A researcher comparing efficiency frontiers could run Muse Ember alongside rivals on benchmark suites like Artificial Analysis's leaderboard to measure intelligence per dollar spent ¹.

A small model for agent chores

Meta called Muse Ember its "smallest and most efficient" entry aimed squarely at the kind of repetitive work AI agents handle all day: making tool calls, running coding subtasks, routing requests, pulling information from documents and parsing large files ².

The company says the model lands right on what it calls the efficiency frontier, meaning no model scores higher for less per task on Artificial Analysis's leaderboard ². Meta also claims it matches Claude Haiku 5.5's best score at roughly a third of the price and beats every setting of GPT-6 Luna .

Artificial Analysis published chart data showing Muse Ember's position against those two rivals across different reasoning levels ¹.

How much it costs

For brief prompts, the standard pricing comes to $0.10 for every one million input tokens and $0.50 for every one million output tokens — matching the listed cost of GPT-6 Luna and Claude Haiku 5.5 ¹. That puts it well below pricier models like the recently announced Claude Fable 5.2, which costs about 80 times more at $8 per million input and $40 per million output tokens .

Things get more complex once prompts stretch past certain length limits. All three budget models support up to one million input tokens, but each raises its price after a threshold ¹:

• GPT-6 Luna doubles to $0.20 / $0.02 / $0.75 beyond 272k tokens
• Muse Ember jumps to $0.30 / $0.012 / $1.50 beyond 200k tokens
• Claude Haiku 5.5 climbs to $0.50 / $0.05 / $2.50 beyond 100k tokens ¹

But comparing raw price-per-token misses something important: these days models use very different numbers of tokens to solve the same problem because of "reasoning tokens," which are extra internal calculations a model generates before giving an answer ¹. That is why Artificial Analysis measures intelligence per dollar spent instead .

Scores on common tests

On Artificial Analysis benchmark results reported by Meta, Muse Ember led several categories at maximum reasoning effort: Agents AutomationBench-AA (37.8 versus 13.9 for Haiku and 20.2 for Luna), Agents' Last Exam (61.6 vs. 57.0 vs. 57.1), Personal Superintelligence Index (46.8 vs. 36.1 vs. 37.0), Coding Terminal-Bench 4.0 (40.9 vs. 39.2 vs. 16.4), Coding Agent Index-AA (46 vs. 37 vs. 31), FrontierCode 2.0-AA (34.9 vs. 27.2 vs. 25.2) and Multimodal GDP.pdf-AA (tied at 22.8 with Luna, ahead of Haiku's 20.8) ².

Simon Willison tested the model himself and generated six sets of images of pelicans riding bicycles using Ember's different reasoning levels — minimal, low, medium, high, xhigh and max ¹. The high level failed to return a full image file, and he noted the visual quality did not clearly track with the reasoning strength . He compared the result to images from Muse Wonder, a model about 40 times more expensive, and found the quality comparable .

Muse Ember is available now through Meta AI's platforms.

Sources

1. Introducing Our Smallest, Most Efficient Model: Muse Ember — simonwillison.net
2. Introducing Our Smallest, Most Efficient Model: Muse Ember — research.meta.ai