Meta releases Muse Ember: a small, cheap model for high-volume AI agent work
Meta's Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens. What it does, how Meta says it compares with Claude Haiku 5.5, and what to check before you switch.
- Author
Umair Mateen- Published
- Reading time
- 4 min
- Category
- AI
Ideas in this post
Tap one for a 5-second answer. Underlined words in the article work the same way.
0 of 4 understood
Meta has added Muse Ember to its Meta Model and its Muse Code coding product. Meta calls it its smallest and most efficient model, built for high-volume agent work such as coding tasks, tool calls and document analysis, and prices it at $0.10 per million and $0.50 per million output tokens for standard-length prompts.
What it is
Ember is the low-cost end of Meta's Muse family, which also includes Muse Spark and the higher-priced Muse Wonder. According to Meta's model page and its developer post:
- Inputs: text, images, audio and video.
- : 1 million tokens (a token is roughly three-quarters of a word).
- Reasoning: Meta describes it as its smallest reasoning model, with adjustable reasoning effort.
- Access: through Meta Model API with an API key from dev.meta.ai, using an OpenAI-compatible client, with the model ID
muse-ember-1.0. It is also included in all Muse Code subscription plans, which Meta lists at $5, $15, $50 and $200 a month.
What it costs
Meta's model page lists two price bands per million tokens:
Quick check · Everyday ↔ real
The restaurant and the API, side by side
Switch between the two and see how each part lines up.
- You, the customer
- The menu
- Your order to the waiter
- The kitchen
- Your dish, or "sold out"
You order from the menu and the waiter brings back your dish, without you ever seeing the kitchen.
| Everyday | Real |
|---|---|
| You, the customer | The client (your app) |
| The menu | The documentation |
| Your order to the waiter | The request |
| The kitchen | The server |
| Your dish, or "sold out" | The response, or an error |
From the explainer What is an API? Explained with a waiter in a restaurant
| Prompt size | Input | Cached input | Output |
|---|---|---|---|
| Up to 200k input tokens | $0.10 | $0.004 | $0.50 |
| Above 200k input tokens | $0.30 | $0.012 | $1.50 |
For comparison, Meta's pricing page lists standard Muse Spark at $1.25 input and $4.25 output per million tokens, and Muse Wonder at $4.00 and $20.00. So Ember costs roughly a tenth of Spark for normal prompts. Long prompts over 200k tokens cost three times as much, so it pays to keep inputs trimmed.
How Meta says it performs
Meta published a table comparing Ember at maximum reasoning effort with Claude Haiku 5.5 and GPT-6 Luna, also at maximum effort. Ember scores higher on most of the agent and automation tests in that table, for example 46 against 37 and 31 on an AI coding-agent index. It is not ahead everywhere: on the FrontierCode 2.0 coding test, Haiku 5.5 scores 34.9 to Ember's 27.2. On Artificial Analysis's overall Intelligence Index, Meta reports Ember and Haiku 5.5 level at 36.
These are Meta's own published numbers. The developer post also includes positive comments from early users including Box, Cline, Oracle, Cloudflare and You.com, but those are vendor-selected testimonials, not independent tests.
Who should care
- Teams running high-volume, repetitive AI jobs: sorting documents, tagging images, extracting fields from forms, or routing support tickets, where the cost per call matters more than top-end reasoning.
- Teams building agents that make many small tool calls in a loop, where a cheap, fast model handles the routine steps.
- Anyone comparing small models such as Claude Haiku or GPT mini-tier models on price, who now has another option to test.
Limits and caveats
- Pricing page lag. At the time of writing, Ember's prices appear on its model page and developer post but not yet on Meta's main pricing and rate limits page, which also lists no rate limits for it. Confirm both in your account before you plan production volumes.
- Not stated: Meta's pages we read do not give a maximum output length, a knowledge cutoff, or whether the weights will be released.
- Benchmarks are vendor-run. Results on your own data and prompts may differ, especially for writing quality and complex reasoning.
- Small models make more mistakes on hard problems. Keep a larger model, or a human check, for decisions that carry real risk.
What to do this week
- Pick one high-volume task you already run on an AI model and list its monthly token use and cost.
- Run 50 to 100 real examples through Ember and your current model side by side, and score the outputs against what a correct answer looks like.
- Work out the cost difference using the price bands above, including whether your prompts go over 200k tokens.
- Check data handling and terms in Meta's developer documentation before sending customer data.
- If quality holds, move the routine steps to the cheap model and keep a stronger model for the hard cases.
If you want help testing small models against your own workflows, with approvals and audit trails built in, The DevDrift does this kind of work.
Sources
You’ve read the news. Now understand it.
- Learn · 5 min · Everyday ↔ realWhat is an API? Explained with a waiter in a restaurantAn API is a set of rules that lets one program ask another for data or a task and get an answer back. Explained with a waiter carrying orders between you and the kitchen.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat are tokens in AI? Explained with word-game tilesTokens are the small pieces of text an AI model reads and writes, and the unit used to measure limits and cost. Explained with a word game whose tiles carry chunks of words.Start in 5 seconds →
- Learn · 5 min · Try the dialWhat is a context window? Explained with a clerk's deskA context window is how much text an AI model can look at in one go: your instructions, documents, the chat so far and its own reply. Explained with a clerk working at a desk of fixed size.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat are AI benchmarks? Explained with a school examAn AI benchmark is a standard test: the same questions and the same marking for every model, so scores can be compared. Explained with a school exam that every student sits.Start in 5 seconds →
