Agent work just got 17x cheaper — what that changes for builders
Meta's new Muse Ember model costs 17 times less per task than its previous small model, which means AI agents that repeat the same jobs all day are now cheap enough for many more businesses to run.
- $0.10 / $0.50
- price per million tokens (input / output)
- 17x
- cheaper per task than Muse Spark 1.3
- 37.8%
- Muse Ember's AutomationBench score
- 1M
- token context window
On November 9, Meta launched a new AI model called Muse Ember. It costs $0.10 per million input tokens and $0.50 per million output tokens — about $0.14 per task on standard benchmarks. Meta's previous small model, Muse Spark 1.3, runs the same task for $2.39. That makes Ember 17 times cheaper per task — a real shift for anyone building AI-powered software.
Muse Ember is not trying to be the most powerful AI around. It is designed for the unglamorous work that agents do in the background all day: calling a tool, routing a request, pulling data from a long document, writing a small piece of code. These jobs happen thousands of times in a live system. The cost adds up fast. At Ember's prices, that stops being the main worry.
The benchmarks back up the claim. On AutomationBench — a test of real business workflows from start to finish — Muse Ember scored 37.8%. Claude Haiku 5.5 scored 13.9% and needs three times as much spend to match Ember's score. GPT-6 Luna came in at 20.2%. Haiku does still lead on a coding benchmark called FrontierCode, so Ember is not the best tool for everything — but for agent loops the value is clear.
The honest takeaway: when the cost of running an agent drops 17 times, more things become worth automating. Tasks that were too expensive to run on a loop — checking inventory, classifying support tickets, summarising documents — look different when the cost is near zero. Ask which jobs could run on a cheap agent in your business today. Muse Ember is available now on Meta's API, Google Cloud, and Oracle Cloud.