# Meta releases Muse Ember, a tiny cheap model for the routine work AI agents do all day

> At $0.10 per million input tokens, Meta says it beats Claude Haiku 5.5 and GPT-6 Luna on intelligence per dollar. Haiku still wins on real coding tasks.

- URL: https://devisdead.ai/news/meta-releases-muse-ember-a-tiny-cheap-model-for-the-routine-work-ai-agents-do-al
- Published: 2026-11-09T20:37:10Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: meta, muse, small models, agents
- Originally reported by The Decoder (https://the-decoder.com/meta-launches-muse-ember-a-small-ai-model-built-for-the-repetitive-work-agents-actually-do/).

Not every agent step needs a frontier model. Meta's new Muse Ember is built for the ones that don't. It's Meta's smallest, most efficient model yet, aimed at tool calls, small coding jobs, routing requests, data extraction and long documents, Matthias Bastian reports for [The Decoder](https://the-decoder.com/meta-launches-muse-ember-a-small-ai-model-built-for-the-repetitive-work-agents-actually-do/).

## The numbers

Meta's pitch is value: it claims Ember beats GPT-6 Luna and Claude Haiku 5.5 on intelligence for the money, whatever reasoning level you pick. On the Artificial Analysis index it scores 36 at about $0.14 per task, against 37 for Meta's larger Muse Spark 1.3 at $2.39 per task. That's nearly the same score for about a seventeenth of the cost.

It handles text, images, audio and video, takes up to a million tokens of context, and you can use it today in Muse Code or through Meta's model API, priced at 10 cents per million tokens in and 50 cents per million out.

## Benchmarks vs. vibes

Ember leads most of Meta's own tests, but not all. When the test is actual coding work, on the FrontierCode 2.0-AA benchmark, Haiku 5.5 comes out ahead, 34.9% against Ember's 27.2%. So for the "small coding tasks" Meta advertises, Haiku may still be the better pick.

## Why it matters

The cheap tier is where agent costs actually live. A typical agent loop is mostly routing, extraction and tool calls, with occasional hard reasoning. Pairing a small model for the routine steps with a big one for the hard ones is how teams are cutting bills, which is exactly the trend [Ramp's spending data](/news/companies-are-using-more-ai-than-ever-and-barely-spending-more-on-it-ramp-data-s) picked up. Ember adds another strong, very cheap option to route to. Benchmark it on your own agent's steps before switching.

**[Scoreboard](/scoreboard) update:** AI +1. Agent grunt work gets a model that costs a fraction of a cent per task.

Stay undead.