TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Meta released Muse Ember on November 9, 2026. Artificial Analysis scored its max setting at 36, matching Claude Haiku 5.5, while reporting a lower average cost per index task; the same evaluation found lower factual accuracy but more refusals and a lower hallucination rate.
Meta released Muse Ember on November 9, and Artificial Analysis scored its max reasoning setting at 36 on the firm’s nine-evaluation Intelligence Index, matching Claude Haiku 5.5 in the comparison. Artificial Analysis reported an average cost of $0.14 per index task for Ember, compared with $0.41 for Haiku, while its separate factual-knowledge test showed lower accuracy and more frequent refusals from Ember.
The reported price advantage is for completing tasks in Artificial Analysis’s index, not a universal measure of what every customer’s workload will cost. The source lists Muse Ember’s token prices at $0.10 per million input tokens and $0.50 per million output tokens, the same listed rates as Haiku 5.5 and GPT-6 Luna. Cached input is listed at $0.004 per million tokens, which the source says is 60% below Haiku’s. Artificial Analysis attributes Ember’s lower index-task cost to cheaper cached input and its use of roughly a third as many output tokens as Haiku on the index.
At max effort, Ember recorded 440 output tokens per second and took 3.2 minutes per index task through Meta’s API, according to the model data accessed November 10. The source also lists a one-million-token context window. Its xhigh setting scored 34 at $0.12 per task, with 423 output tokens per second and 2.8 minutes per task. Those results describe the two tested settings; they do not establish performance or cost on other tasks.
Results varied by evaluation. Ember scored 68.1% on AutomationBench-AA, equal to Meta’s more expensive Muse Wonder and above Haiku’s 35.4% in the cited data. Ember did better than Haiku on Terminal-Bench 4.0, 40.9% to 32.8%, but scored lower on FrontierCode 2.0, 27.2% to 34.9%. On the AA-Briefcase knowledge-work test, Ember registered 1,521 Elo against Haiku’s 1,574; on GDP.pdf, Ember scored 22.8% against 20.8%.
Muse Ember’s score of 36 comes with a trade-off
Meta’s model matched Claude Haiku 5.5 on Artificial Analysis’s Intelligence Index at a lower average cost per index task. Its factual test results show more refusals and lower accuracy.
The index cost advantage
Artificial Analysis measured these average costs on its own mix of nine evaluations.
Factual answers involve a trade-off
On AA-Omniscience, Ember declined more often and answered correctly less often than Haiku.
| Measure | Muse Ember | Claude Haiku 5.5 |
|---|---|---|
| Correct answers | 21.1% | 34.7% |
| Hallucination rate | 17.4% | 34.2% |
| Refusals | 64% | Not provided |
Results vary by task
Ember led Haiku on some cited evaluations and trailed on others.
- AutomationBench-AA: 68.1% vs 35.4%
- Terminal-Bench 4.0: 40.9% vs 32.8%
- GDP.pdf: 22.8% vs 20.8%
- FrontierCode 2.0: 27.2% vs 34.9%
- AA-Briefcase: 1,521 Elo vs 1,574
Check your own workload
The benchmark cost may not carry over when prompts, cached context and output length change.
Use prompts and workflows your team actually runs.
Compare accuracy, refusals, speed and total cost against your current models.
Decide where the price and reliability trade-off fits your error costs.
A Lower Cost for Some Workflows
The comparison points to a possible cost advantage for organizations running many repeated tasks, such as routing, extraction, tagging and form filling. Artificial Analysis’s index result suggests Ember may complete its mix of tasks for less than Haiku 5.5, but the source itself cautions that an index is not a company’s own workload. The observed price depends on the benchmark’s prompts, cached context and output length.
Ember’s refusal pattern also affects where it may fit. In the AA-Omniscience test, it answered correctly on 21.1% of questions, declined 64%, and had a reported hallucination rate of 17.4%. Haiku answered 34.7% correctly and had a 34.2% hallucination rate in that test. These are benchmark measurements, not a guarantee that Ember will avoid unsupported answers in business use. For factual work, a refusal can still interrupt a process and require a person or another system to respond.
The figures make the model’s trade-off relevant to buyers weighing cost, task performance and factual reliability. A lower benchmark cost may matter at high volume, while weaker factual results may make it unsuitable for tasks that depend on answering obscure questions. Teams would need to assess it on their own prompts and error costs.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
As an affiliate, we earn on qualifying purchases.
How Ember Compares on the Index
Muse Ember is described in the source as the efficient tier of Meta’s Muse model line, a proprietary model available through Meta’s API and in Muse Code. It accepts text and images and produces text. Artificial Analysis tested it at max and xhigh reasoning effort. Its index is described as covering nine evaluations, and the source reports scores and average API costs for those tested configurations.
Within the comparison presented, Claude Haiku 5.5 also scored 36 at max effort, for $0.41 per index task. Meta’s Muse Wonder scored 44 at $6.95 per task. The source also lists Claude Fable 5.2 at its highest effort with a score of 65 and a task cost of $26.54. These figures illustrate the spread in the cited data, but they do not show that the models are interchangeable or that a given score predicts performance on a specific company’s work.
Artificial Analysis’s coding-agent index gave Muse Ember 46 in Meta’s Muse Code, compared with 37 for Haiku 5.5 in Claude Code and 31 for GPT-6 Luna in Codex. This is a comparison across models used in different coding tools, as presented by Artificial Analysis. It sits alongside Ember’s lower FrontierCode 2.0 result, showing that performance can differ across evaluations.
Limits of the Benchmark Evidence
The cited results do not show how Muse Ember performs across real customer workloads, or whether the reported task-cost difference holds when prompts, context lengths and usage patterns change. The source recommends measuring the model on a buyer’s own tasks. It does not provide a broader independent deployment study or results over time.
The AA-Omniscience figures also leave open how refusal behavior varies across subject areas and how often users can resolve a refusal with added context. The source reports overall rates, but does not establish whether the model’s factual weakness is concentrated in particular fields. It is also unclear from the supplied material how the listed benchmark results will change with later model updates.
Testing Ember on Company Tasks
Organizations considering Muse Ember can compare it with their current models using representative prompts and their own measures for accuracy, refusal rates, speed and total cost. That would help determine whether the lower cost reported on Artificial Analysis’s index applies to their workloads and whether the model’s performance is adequate for each task.
The source identifies access through Meta’s API and Muse Code, but provides no further announcement schedule or follow-up benchmark date. Further comparisons or customer testing may clarify where Ember’s cost and refusal profile is useful. Until then, the published results describe a particular set of evaluations and tested settings.
Key Questions
What did Meta announce?
Meta released Muse Ember on November 9, 2026. The source describes it as a proprietary model available through Meta’s API and Muse Code.
How did Muse Ember score against Claude Haiku 5.5?
Artificial Analysis gave both models a score of 36 on its Intelligence Index at max effort. It reported an average cost of $0.14 per index task for Ember and $0.41 for Haiku.
Does the lower task cost mean Ember is cheaper for every use?
No such conclusion follows from the benchmark alone. The reported figure is an average cost for tasks in Artificial Analysis’s index; cost on a company’s work can depend on its prompts, cached input and output length.
Why did the source say Ember may guess less?
On AA-Omniscience, the model declined 64% of questions. Artificial Analysis reported a 17.4% hallucination rate for Ember in that test, alongside 21.1% accuracy. The results are specific to that evaluation.
Source: Thorsten Meyer AI
