Meta released Muse Ember on 9 November, the smallest and cheapest model in its Muse family. Its API price is $0.10 per million input tokens and $0.50 per million output tokens. That is exactly what Claude Haiku 5.5 and GPT-6 Luna charge, so the price list cannot tell you which of the three to buy. This article covers what Meta shipped, what the independent tester Artificial Analysis measured, what a finished job costs on each model, and where each one fits. The short version: Ember ties Haiku 5.5 on the intelligence score at about a third of the cost per task, but it is not better at everything, and Meta has not yet published its data terms for the model.
What Meta released on 9 November
According to Meta's own announcement, Muse Ember is built for the high-volume, repeatable steps inside an agent workflow, not for the hardest problems. Meta's line is that Muse Wonder plans and Muse Ember handles the volume. The facts that matter to a buyer:
- Where you can call it. Through the Meta Model API as
muse-ember-1.0and inside Muse Code. Meta also says Ember, Spark and Wonder are now available on Google Cloud and Oracle Cloud. - Context window. 1 million tokens in, up to 512,000 tokens out.
- Inputs. Text, images and audio, according to Meta.
- No downloadable weights. Meta keeps the model private, so this is not an open-weight release like the models we covered in open-weight AI models and what they cost.
The price: three models, one price list
From Meta's announcement, Anthropic's pricing documentation and OpenAI's API pricing page, all read on 12 November 2026:
| Model | Input (per 1M tokens) | Cached input | Output (per 1M tokens) |
|---|---|---|---|
| Meta Muse Ember | $0.10 | $0.004 | $0.50 |
| Claude Haiku 5.5 | $0.10 | $0.01 | $0.50 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
The identical headline price hides three differences that show up on a real bill:
- Long prompts cost more, at different thresholds. Meta's page says Ember charges $0.30 input and $1.50 output once a prompt passes 200,000 tokens. Anthropic's documentation says Haiku 5.5 charges $0.50 input and $2.50 output once a prompt passes 100,000 tokens, and that the count includes cached tokens. If your agent drags a big document or a long chat history into every call, the cheaper model on paper may not be the cheaper one on your invoice.
- Cached input is cheapest on Ember. $0.004 per million against $0.01 on the other two, which matters if the same long instructions are re-sent on every call.
- Luna has its own long-context rate. OpenAI's pricing page lists separate long-context, Batch and Fast rates for Luna. We have not reproduced them here because the page does not make the comparison simple. Check it against your own prompt sizes.
What Artificial Analysis measured
Artificial Analysis runs the same set of tasks on every model and reports a score, a price and the cost of an average task. Its pages for the three models show:
Ember and Haiku 5.5 both score 36, and Ember finishes an average task for $0.14 against Haiku's $0.41, about three times less (our division of Artificial Analysis's figures). Luna is nearly as cheap at $0.15 but scores 28, eight points lower. Artificial Analysis lists Ember at 448 output tokens per second and 3.1 minutes per task, and Haiku at 255 tokens per second. Meta's own page quotes the same score and cost figures, and attributes them to Artificial Analysis dated 9 November.
One honest note on the number: it is a score for the "Max" reasoning setting, which spends the most tokens. Your cost on easier work, at lower reasoning settings, will differ, and so will the gap.
Where Ember wins, and where it does not
Meta's launch table compares Ember with Haiku 5.5 and Luna. Several of those tests are Meta's own, so we leave them out. The ones Meta attributes to Artificial Analysis or to a public benchmark:
- Business-task automation. On AutomationBench-AA, which tests multi-step work across business apps, Ember scores 37.8%, Luna 20.2% and Haiku 13.9%. This is the closest benchmark to what a small business actually asks an agent to do.
- Computer use. On OSWorld 2.1 (partial credit), Ember 69.5% and Haiku 68.9% are close, and Luna is at 50.5%. On the stricter binary scoring it is 33.8, 27.8 and 17.0.
- Coding agents. Ember 46 on Artificial Analysis's Coding Agent Index against Haiku's 37 and Luna's 31.
- Where Haiku wins. On FrontierCode 2.0, Haiku scores 34.9% and Ember 27.2%. If your work is hard single-shot coding, Ember is not the pick.
- A caution on Terminal-Bench. Meta reports 40.9% for Ember against 39.2% and 16.4%, but each model was run in the best harness published for it, so the test setups differ. We do not treat it as a like-for-like result.
When three models share a price list, the price stops being the question. Tokens used per finished job, and what breaks when the prompt gets long, are what decide the bill.
What 1,000 jobs cost, and why the table above matters more
Take an illustrative job: a support agent that reads 5,000 tokens and writes 1,000. That is our assumption, not a measured workload. Run 1,000 of them and you send 5 million tokens in and get 1 million out. At $0.10 and $0.50 per million that is $0.50 for input plus $0.50 for output, $1.00 on all three models, because the rates match.
The real difference arrives when the agent has to work, retry and check its own steps. A reasoning model can spend many times more tokens thinking than answering, and that is exactly what Artificial Analysis's cost-per-task figure captures: it multiplies the price by the tokens each model actually used. That is why the same price list produced $0.14 for Ember and $0.41 for Haiku. We made the same point in our guide to the cheapest AI model by cost per task, and Ember is a clean example of it.
Who should switch, and who should wait
- Switch to test it: teams running high-volume agent steps, such as sorting incoming messages, filling forms across apps, extracting fields from documents or drafting routine replies. If you already use Haiku 5.5 for this, run Ember on the same 50 real tasks and compare cost and error rate on your own work.
- Stay on Haiku 5.5: hard coding tasks where its FrontierCode lead matters, or where you already depend on Anthropic's tooling and terms.
- Be careful with Luna: it is as cheap as Ember per task but scored eight points lower on the same index. It suits simple, high-volume chat where the quality bar is modest.
- Wait, if you handle customer data: Meta's Ember announcement does not state its data terms, meaning whether prompts are used for training, how long they are kept, or where they are processed. We checked Meta's launch page and developer front page and found nothing on it. Meta has had separate data modes on earlier Muse models, so ask before you send customer records.
The same rule applies to any model that changes by a few cents per job: the saving is real only if the cheaper model also gets the job right the first time. A model that is a third of the price and wrong twice as often is not cheaper.
What this means for small-business AI, and where to start
The useful change is that a score of 36 on Artificial Analysis's index now costs about fourteen cents per average task on this model. At that level the model bill is rarely the biggest cost of an agent; the work of connecting it to your accounts, your phones and your stock is. That is the part we build: the IO Snack AI Assistant and IO Snack Accounts are where this kind of cheap, high-volume model work ends up.
If you want the broader picture before choosing, read our comparison of Claude Haiku 5.5's price and benchmarks, and how Meta's bigger model compares in Muse Wonder: price, benchmarks and who should use it. If you would like help testing models on your own workload, talk to us.
