Meta's Muse Ember, the AI for agents: a third of Claude Haiku's cost per task
Muse Ember costs the same per token as Claude Haiku 5.5, but almost three times less per task. The full math, use cases and limits.
Muse Ember is Meta's smallest and most efficient model, introduced on November 9, 2026 for the repetitive work of AI agents. It costs the same per token as Claude Haiku 5.5 ($0.10 per million input tokens and $0.50 per million output), but according to Artificial Analysis it comes out almost three times cheaper per task: $0.14 versus $0.41.
What changed with Muse Ember

The news isn't the list price. Meta built Muse Ember for the daily work of agents: calling tools in sequence, running tests, making small code fixes and extracting data from documents. It is available in the Meta Model API (identifier muse-ember-1.0), on OpenRouter, on Google Cloud Marketplace and in Muse Code. It is not in the Meta AI app or the Muse agent. It is also proprietary: Meta did not release the weights.
It costs $0.10 per million input tokens and $0.50 per million output tokens on prompts up to 200K tokens; above that, $0.30 and $1.50. Context is 1 million tokens and output can reach 512K.
Per token vs. per task
A model doesn't charge per task: it charges per token, per little piece of text it reads and writes. Two models with the same price per token can cost very different amounts if one writes three times more to reach the same result. Artificial Analysis, an independent firm, measured exactly that: Muse Ember wrote about a third of what Haiku 5.5 wrote (about 115 million output tokens versus 298 million in the index run).
That matters more for an agent. A chat answers once; an agent takes many steps, and every step is a call that rereads what came before. A chatty model multiplies that habit by every step and every task in the month.
The Muse Ember math: 10,000 tasks a month

This is our own analysis using Artificial Analysis figures. Both models score 36 on the intelligence index, so the comparison is fair.
- 1,000 tasks: $140 with Muse Ember versus $410 with Haiku 5.5. A difference of $270.
- 10,000 tasks a month: $1,400 versus $4,100, a gap of $2,700 a month.
- Over a year that is $32,400.
- 41 divided by 14 is almost three times.
Three business cases (our own analysis)
- Accounting firm. With 800 invoices and receipts a month, Muse Ember pulls vendor, date, total and taxes into a spreadsheet, and one person checks one in ten.
- Online store. With 3,000 messages a month, it classifies each one (shipping, returns, invoices), routes it and leaves a draft. Ask it for a single label from a closed list, not a paragraph, because every extra word is output you pay for.
- Agency or developer. A big model plans and Muse Ember executes the subtasks. Muse Wonder costs $6.95 per task, about 50 times what Muse Ember costs: in an illustration with 20 tasks that is $139 versus $2.80.
Who it competes with and who it's not for
Artificial Analysis also measured DeepSeek V4.1 Flash (index 33, $0.60 per task) and GLM-5.3-Flash (index 28, $0.58). Muse Ember scores 36 at $0.14, but it is proprietary and lives only in the cloud of Meta and its partners.
It isn't for anyone making few requests: at 100 tasks a month the saving is $27. Nor for anyone who needs code someone else will merge without review: on FrontierCode, Haiku 5.5 wins, 34.9 to 27.2.
What Meta isn't telling you
The cost per task and the score of 36 were measured by Artificial Analysis, not Meta. Meta, on the other hand, ran four rows of its comparisons on its own, and the Terminal-Bench table mixes sources. On the knowledge test, Muse Ember gets 21.1% right versus 34.7% for Haiku 5.5: it refuses to answer about 64% of questions and hallucinates half as much as Haiku (17.4%). It is good for extracting data from a document you give it, and bad for asking things from memory.
What breaks in practice:
- The 200K-token step: a 300K-token prompt costs $0.09 for input alone, at triple the per-token rate.
- Audio and video: Meta says it accepts them, but Artificial Analysis only lists text and image.
- Unreviewed code: you'll see failing tests.
- Access: only through the API and Muse Code, not the app.
How to apply it in your business?
These are the five questions to decide:
- How many tasks do you really do per month, counted? Under a thousand, and the difference doesn't reach $270.
- Do you extract, classify or execute? If so, keep going. If it's code to merge, stick with Haiku.
- Do your prompts go past 200K tokens? Split them.
- Do you give it the document or ask it to remember? Give it the document.
- Test 50 of your own cases with two models and measure the output tokens.
And one warning: don't compare list price, compare cost per task. For another example of cost per task between models, read the full breakdown of Claude Fable 5.2.
Frequently asked questions
What is Muse Ember?
It is Meta's smallest and most efficient model, built for the repetitive work of agents. It was introduced on November 9, 2026.
How much does Muse Ember cost?
$0.10 per million input tokens and $0.50 per million output tokens on prompts up to 200K tokens. Above that, $0.30 and $1.50.
Is Muse Ember cheaper than Claude Haiku 5.5?
Per token it costs the same. Per task, according to Artificial Analysis, it costs $0.14 versus $0.41, because it writes about a third of the output tokens.
Where can you use Muse Ember?
In the Meta Model API, OpenRouter, Google Cloud Marketplace and Muse Code. It is not in the Meta AI app.
Is it open source?
No. Muse Ember is proprietary: Meta did not release the weights.
Conclusion
With Muse Ember, price per token is no longer the number that matters: what matters is how many tokens the model burns to finish your job. The question that remains is whether it holds up on your tasks, not the ones on the chart. Test it on your own cases before switching anything.