Meta's Muse Ember prices agent work at $0.10 input

Agent work just got cheaper again.

Meta's new Muse Ember costs $0.10 per million input tokens, and Meta says it beats the small models from Anthropic and OpenAI on the jobs agents repeat all day.

Meta's smallest Muse is priced for volume

Meta released Muse Ember on November 9. It calls Ember its smallest and most efficient model, built for tool calls, coding subtasks, routing, extraction and long document parsing.

The price for prompts up to 200,000 tokens, per million tokens:

$0.10 input, $0.50 output.

Cached input costs $0.004 per million. Past 200,000 tokens the prices triple, to $0.30 input and $1.50 output. The context window is 1 million tokens, with up to 512,000 output tokens.

Ember runs on the Meta Model API and in Muse Code as muse-ember-1.0, and Meta says it is also on Google Cloud and Oracle Cloud. It takes text, images, audio and video, and supports the Responses API. Meta's post does not say the weights will be released.

The small model fight is now about agents

Yes, every lab says its small model wins. Meta's chart at least shows where Ember loses.

On Terminal-Bench 4.0, Meta reports Ember at 40.9%, against 39.2% for Claude Haiku 5.5 and 16.4% for GPT-6 Luna. On AutomationBench-AA it reports 37.8%, against 20.2% for GPT-6 Luna and 13.9% for Haiku 5.5.

The losses are in the same table. On FrontierCode 2.0-AA, Meta lists Haiku 5.5 at 34.9% and Ember at 27.2%. On OSWorld 2.1 with partial credit, the two are close: 69.5% for Ember, 68.9% for Haiku 5.5.

All of these figures are Meta's own.

The cost line is the one to read. Meta says an Artificial Analysis task costs about $0.14 on Ember, against $2.39 on its larger Muse Spark 1.3. On that index, Meta lists Ember at 36 and Spark 1.3 at 37.

One point of score for about a seventeenth of the price.

Where a ten cent model fits

If you run AI agents, most of your calls are not hard. They route a request, pull a field from a document, call a tool, format a reply.

Those calls are where a $0.10 model pays for itself. Keep a frontier model for planning and judgment, and hand the repetitive steps to a small one.

Two things to check before you switch. Long document jobs past 200,000 tokens pay three times the base price. And Ember is only on Meta's API and the two clouds Meta named, so a move means a new vendor in your stack.

Meta has been more open about this launch than its last big one. When it shipped Muse Wonder in October, it gave no benchmarks or prices. Ember came with both.

Meta also says it evaluated Ember under its Meta Superintelligence Scaling Framework, and that the model stayed below the high-risk thresholds for the categories it assessed.

Watch for scores Meta didn't write

In Meta's post, launch partner Cloudflare says Ember "ran up to 30% faster at the tail" than a comparable model. That is a partner talking at a launch.

The number to watch is an independent one. Until outside tests line up with Meta's chart, treat it as a pitch.

Run your ten most common agent calls through Ember. Then compare the bill.