Meta Muse Ember: $0.10 per Million Tokens for Repeat Agent Work
In short
- Meta announced Muse Ember on November 9, its smallest and cheapest model, built for agents that repeat the same work all day.
- Price: $0.10 per million input tokens and $0.50 per million output tokens, with a 1M token context window.
- The money angle: one big agent run costs about 20 cents instead of $16 on a frontier model. The quality gap is real, though.
Meta just released Muse Ember. How much does it cost? Ten cents per million input tokens. That is the headline, and it is why you should look at it.
What happened
Muse Ember (model ID muse-ember-1.0) is Meta’s smallest, most efficient model. Meta says it is made for the repeat work of agents: tool calls, coding subtasks, routing, data extraction and parsing long documents. It handles text, images, audio and video, with a 1M token context and up to 512K tokens of output. Meta did not publish a parameter count.
Pricing for prompts up to 200K tokens: $0.10 input, $0.004 cached input, $0.50 output, all per million tokens. Above 200K tokens it is $0.30, $0.012 and $1.50. You can use it in the Meta Model API, in Muse Code, and on Google Cloud and Oracle Cloud. Meta also has a test page at dev.meta.ai.
Meta compared it with Claude Haiku 5.5 and GPT-6 Luna at max reasoning. On the Artificial Analysis Intelligence Index it ties Haiku at 36 (Luna: 28). On AutomationBench-AA it scores 37.8% against 13.9% for Haiku and 20.2% for Luna. But on FrontierCode 2.0-AA, Haiku wins: 34.9% against 27.2%. These are Meta’s numbers.
How to use it
Use it for the boring, repeated steps in an automation: sort incoming emails, pull fields out of invoices, route support tickets, summarise long PDFs. Keep the expensive model for planning and hard decisions. Cached input at $0.004 per million helps when the same long prompt runs again and again.
Do the math: one run with 1M input and 200K output tokens costs $0.10 + $0.10 = $0.20 on Muse Ember, about €0.17. The same run on Claude Fable 5.2 costs $16, about €14. Do 1,000 such runs a month: about €170 against about €13,900. Of course Fable is a far stronger model, so this is a ceiling on savings, not a fair swap. Still, if a task is simple, you should not pay for the heavy model.
What it means
Meta is going after the cheap end of the market, where Haiku and GPT-6 Luna sit. Partners like Cloudflare, Box and You.com tested it. Cloudflare reports up to 30% faster tail latency against a comparable model. Nothing on the page says what license it has, so do not assume it is open.
Ivan’s Take
Worth a test for anything repetitive. The price is hard to ignore, and it beats Haiku on the automation benchmark Meta chose. I would not trust it with hard coding, since Haiku is ahead there. Run your own 50 real tasks before you move anything.
Source: Meta
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
