Meta released Muse Ember on 9 November, a small model for the repetitive steps AI agents take all day, such as tool calls and data extraction. It costs $0.10 per million input tokens and $0.50 per million output tokens, the same list price as Claude Haiku 5.5 and GPT-6 Luna. Meta’s case rests on cost per task: it says Ember matches the best benchmark score of Haiku 5.5 at about a third of the cost.
Meta calls Ember its smallest and most efficient model. It is built to work under a larger model, or in Meta’s words: “Muse Wonder plans, Muse Ember handles the volume.” Meta names tool calls, coding subtasks, routing, extraction and long documents as its main jobs. Ember reads text, images, audio, video and PDFs, and answers in text.
- Model ID: muse-ember-1.0
- Where: Muse Code and the Meta Model API, plus Google Cloud and Oracle Cloud. OpenRouter also lists it, with Meta as the only provider.
- Context window (the text it reads at once): 1 million tokens, with answers of up to 512,000 tokens
- Weights: Meta offers Ember only as a hosted model and does not mention a download
For organisations that build agents, small models like this decide most of the bill. In a typical agent workflow, most model calls are small steps: sorting a ticket, choosing a tool, reading an amount from an invoice. Those steps can run thousands of times a day. Ember adds a third option at the same list price. The choice now rests on what a finished task costs and how often the model gets a step wrong.
Ember’s saving per task comes from tokens, caching and long prompts, not the list price
The three models charge the same for short prompts and part ways on caching, long prompts and the cost of a finished task.
| What is measured | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| Input, per million tokens | $0.10 | $0.10 | $0.10 |
| Output, per million tokens | $0.50 | $0.50 | $0.50 |
| Cached input, per million tokens | $0.004 | $0.01 | $0.01 |
| Higher rates start above | 200,000 tokens | 100,000 tokens | 272,000 tokens |
| Input and output above that | $0.30 and $1.50 | $0.50 and $2.50 | $0.20 and $0.75 |
| Artificial Analysis Intelligence Index | 36 | 36 | 28 |
| Cost per index task | $0.14 | $0.41 | $0.15 |
| AutomationBench-AA | 37.8% | 13.9% | 20.2% |
| Coding Agent Index-AA | 46 | 37 | 31 |
Prices come from Meta, Anthropic and OpenAI. The scores and costs per task are Artificial Analysis runs at maximum reasoning effort that Meta published, and Meta labels Ember’s own costs as estimates.
Because the list rates match, Ember’s lower cost per task has three possible sources. It may need fewer tokens to finish a task. Its cached input is cheaper, which helps agents that resend the same instructions on every call. And its long prompts cost less. Haiku 5.5 charges five times its base rate once a prompt passes 100,000 tokens. Ember triples its rate, and only above 200,000. Meta does not say which of these explains its chart.
Against GPT-6 Luna the picture changes. Ember costs slightly less per task and scores eight points higher on the index.
Ember leads Haiku 5.5 on Zapier’s automation tasks and trails it on FrontierCode
The independent numbers are the strongest part of Meta’s case. Artificial Analysis runs AutomationBench on a private set of 657 Zapier tasks: multistep work in business software, the kind agents do. Ember completed 37.8% of them, against 13.9% for Haiku 5.5. On the coding test FrontierCode 2.0, Haiku 5.5 leads with 34.9% to Ember’s 27.2%.
Read the rest of Meta’s results with more care. Four of the tests are Meta’s own runs, and Meta writes that it may not have tuned the rival models optimally. On Terminal-Bench 4.0, each model ran in a different tool, and two of the three scores come from other publishers. Partner speed claims, such as Cloudflare’s report that Ember was up to 30% faster on its slowest responses, do not name the comparison model.
Meta rules out training on Ember prompts and names no data region
Meta’s developer documentation says Ember runs only in Standard mode, where your content is not used to train Meta models. It does not say where prompts are processed or how long Meta keeps them. The announcement does not mention Europe. OpenRouter states that Ember is not available through its EU in-region routing, which keeps requests on EU servers, and Google Cloud’s Meta page lists no regions.
That gap matters as soon as a prompt contains personal data, such as a customer email or an invoice with a name on it. Under the GDPR, you need to know where that data goes and which safeguards cover a transfer outside the EU.
Test Ember on your own work before you move anything. Run a hundred real subtasks through Ember and your current small model. Compare the cost per completed task and the error rate, not the price per token. Teams on Haiku 5.5 whose prompts often pass 100,000 tokens have the most to gain. Keep personal data out of those tests until Meta says where Ember processes it.
For more information, visit the official announcement of Muse Ember on the Meta blog.
