Meta has released Muse Ember, the smallest and cheapest model in its Muse family. It is not built to chat. Meta designed it for the repetitive jobs AI agents run all day, such as tool calls, routing, data extraction and small coding fixes.
The pitch is aimed squarely at Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT 6 Luna, the two budget models it compares itself with in its launch post. Meta’s claim is simple: the same results for a fraction of the running cost.
What Muse Ember is
In its launch post, Meta calls Muse Ember its smallest and most efficient model for complex reasoning. It reads text, images, audio and video, can operate a computer from screenshots, and has a one million token context window with up to 512,000 tokens of output.
Meta suggests pairing it with Muse Wonder, its top model: Wonder plans the work and Ember handles the volume. Box, Cline, Oracle and Cloudflare are among the early testers quoted in the post.
How the price compares
Muse Ember costs $0.15 per million input tokens and $0.73 per million output tokens. On short prompts that is the same list price as Claude Haiku 5.5. The gap opens elsewhere: Haiku 5.5 charges five times more for prompts over 100,000 tokens, and Ember’s cached input is cheaper at well under one cent per million tokens.
Meta bills in US dollars, so these figures move with the exchange rate. Its bigger model, Muse Spark 1.3, costs $1.81 in and $6.17 out per million tokens.

The cost per task claim
The headline number is cost per task on the Artificial Analysis index, an independent test suite. Meta puts Ember at about $0.20 per task and says it matches Haiku 5.5’s best score at about a third of the cost. On Meta’s chart, Haiku 5.5 at that score costs about $0.60 per task.
Both score 36 on the index, so the saving comes from efficiency, not extra intelligence. Ember wins clearly on the AutomationBench test, at 37.8 per cent against 13.9 per cent for Haiku 5.5, but trails it on the FrontierCode coding test, 27.2 per cent against 34.9 per cent. Meta notes some cost figures are estimates.
Why Australian businesses should care
Most of what a business agent does is cheap, repetitive work: reading an invoice, filing a ticket, checking a record. That is where model prices add up, and where a cheaper small model can change whether an automation project pays for itself.
Availability looks straightforward but is not spelt out. Meta’s help centre says its Model API works in most countries where Meta AI is available and billing is supported, and Australia is not on its list of restricted territories. Meta does not publish a full list of supported countries.
Test before you switch
A model that is cheap on a benchmark can still cost more if it needs extra retries on your own data. The sensible move is to run Ember next to your current model on a sample of real tasks and compare the bill and the error rate.
Security is worth checking too, because agents read files and web pages they did not write. Meta says it tested whether Ember resists hidden instructions in web pages, documents and tool outputs, and whether it asks a person before high impact actions. It also says the model stayed below the high risk thresholds of its safety framework.
If your stack already runs on Claude, our earlier report on Claude Haiku 5.5 sets out what Ember is being measured against.
