🔍 Read the full analysis: Muse Ember’s 36 Score Comes With An Uncertainty Caveat on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Artificial Analysis reports that Meta’s Muse Ember (max) scored 36 on its Intelligence Index, matching Claude Haiku 5.5, while costing $0.14 per index task versus $0.41. The comparison comes with a caveat: Muse Ember answered only 21.1% of questions correctly on the AA-Omniscience test, while declining 64%; its reported hallucination rate was lower than Haiku’s.
Meta’s Muse Ember (max) scored 36 on Artificial Analysis’ Intelligence Index, matching Claude Haiku 5.5 while costing $0.14 per benchmark task compared with Haiku’s $0.41, according to results published November 9 and 10, 2026. The lower-cost result comes with a marked limitation: on the index’s factual-knowledge test, Muse Ember answered correctly less often and declined many questions rather than guessing.
Artificial Analysis describes its Intelligence Index as a test spanning nine evaluations. Its model pages list Muse Ember (max) at 36 and $0.14 per index task, while Claude Haiku 5.5 (max) also scores 36 at $0.41. The benchmark’s per-task cost is an average API cost for completing one index task, not a price guarantee for every customer’s workload. The organization calls Muse Ember the cheapest model scoring 36 or higher in its comparison; that characterization applies to the models and results it evaluated.
The reported token prices are $0.10 per million input tokens and $0.50 per million output tokens, matching the listed rates for Haiku 5.5 and GPT-6 Luna. Artificial Analysis attributes Muse Ember’s lower task cost partly to cached input priced at $0.004, which it says is 60% cheaper than Haiku’s cached input, and to the model using roughly a third as many output tokens as Haiku to complete the index. The benchmark reports 440 output tokens per second and 3.2 minutes per index task for Muse Ember (max); its context window is listed at one million tokens.
The results are not uniformly strong. Muse Ember scored 68.1% on AutomationBench-AA, matching Meta’s more expensive Muse Wonder and exceeding Haiku 5.5’s reported 35.4%. It beat Haiku on Terminal-Bench 4.0, 40.9% to 32.8%, but trailed on FrontierCode 2.0, 27.2% to 34.9%. Artificial Analysis also reported a score of 46 for Muse Ember in Meta’s Muse Code on its Coding Agent Index, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex. These are benchmark results, not evidence that the model will perform the same way on every coding or business task.
Muse Ember’s 36 Score Comes With a Caveat
Muse Ember (max) matched Claude Haiku 5.5 on Artificial Analysis’ Intelligence Index at a lower benchmark-task cost. Its factual-question results show why the score is not a complete measure of reliability.
The benchmark cost comparison
The reported average is for completing an index task, not a guaranteed price for every workload.
| Model | Index score | Cost per task |
|---|---|---|
| Muse Ember (max) | 36 | $0.14 |
| Claude Haiku 5.5 (max) | 36 | $0.41 |
| Muse Ember (xhigh) | 34 | $0.12 |
Factual answers need a closer look
On AA-Omniscience, accuracy and refusal behavior tell different parts of the story.
Results vary by task
The index tie does not mean the models perform alike across every evaluation.
- AutomationBench-AA: 68.1%, matching Muse Wonder and exceeding Haiku’s 35.4%.
- Terminal-Bench 4.0: 40.9%, ahead of Haiku’s 32.8%.
- AA-Omniscience accuracy: 34.7%, compared with Muse Ember’s 21.1%.
- FrontierCode 2.0: 34.9%, compared with Muse Ember’s 27.2%.
Test before changing model routes
Evaluate the model on representative work, including what happens when it declines.
Compare models on representative prompts and company data.
Track accuracy, refusals, latency, and total API cost.
Record whether a human or another model must handle unanswered requests.
Lower Cost, With a Knowledge Trade-Off
The result gives businesses a reason to test Muse Ember for high-volume, repeatable workflows, where per-task costs and repeated context can add up. Tasks such as extraction, tagging, form completion, and steps in an agent workflow may rely on information supplied in documents or retrieved from a company’s systems, rather than facts the model must recall unaided. Artificial Analysis’ results suggest the model may be competitive on some such tasks, but they do not establish savings or quality for a particular company. Local evaluations are needed before changing model routes or budgets.
The factual-question results change how that price comparison should be read. On AA-Omniscience, Muse Ember’s 21.1% accuracy was below Haiku 5.5’s 34.7% and Gemini 3.8 Flash (high)’s 50.5%. It declined 64% of the questions, and Artificial Analysis reported a 17.4% hallucination rate, compared with 34.2% for Haiku. A refusal can reduce unsupported answers, but it also means the system may fail to complete a task without a fallback. Buyers should weigh both answer quality and abstention behavior rather than treating a lower hallucination rate as a general measure of reliability.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Samsung Galaxy Tab A9+ 11-Inch Android Tablet, 64GB, Graphite
DJI Osmo Mobile 8 3-Axis Phone Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
How the Benchmark Costs Compare
Meta released Muse Ember on November 9, 2026, as the efficient tier of its Muse model line. The proprietary model is available through Meta’s API and in Muse Code, Meta’s coding agent. It accepts text and images and produces text. Artificial Analysis tested two reasoning-effort settings: max and xhigh. The xhigh version scored 34 and cost $0.12 per index task, according to model data accessed November 10; the headline comparison is for the max setting, which scored 36.
The cost difference is specific to the benchmark’s usage pattern. Although Muse Ember and Haiku 5.5 have the same listed input and output token rates, Artificial Analysis says cached input is cheaper for Muse Ember and its benchmark run uses fewer output tokens. The organization’s comparison also places Meta’s Muse Wonder at 44 for $6.95 per task and Anthropic’s Claude Fable 5.2 at 65 for $26.54 at its highest effort level. Those figures show that higher scores in this dataset can come at much higher costs, but they do not establish a universal price-performance ranking outside the tested models and tasks.
What the Score Does Not Show
The published figures are benchmark measurements; they do not show how Muse Ember will perform on a reader’s specific prompts, data, or production traffic. The source material does not provide independent replication of these results or enough detail here to determine how sensitive the comparisons are to prompt design, model updates, or changes in API pricing. The one-million-token context window and reported speed also do not establish that every request will run at that rate or cost.
It is also unclear how often users will encounter a refusal in ordinary applications. The 64% decline rate is from AA-Omniscience, a test of difficult factual questions, and should not be treated as a refusal rate across all tasks. The cited summary does not specify how the model’s behavior changes when it receives retrieved evidence, how consistently a system can distinguish a justified refusal from an avoidable one, or what the downstream cost of fallback calls would be.
Test Workloads Before Switching
Organizations considering Muse Ember can compare it with their existing models using representative tasks, measuring accuracy, refusal frequency, latency, and total API cost rather than relying on the index score alone. For knowledge-intensive work, testing with supplied documents or retrieved company data may show whether grounding helps address the weakness reported on AA-Omniscience. A practical evaluation should also record what happens when the model declines, including whether a human review or another model can handle the request.
Meta’s model is available through its API and Muse Code, while Artificial Analysis’ pages provide the cited benchmark and pricing comparisons. The source material does not state a further release date or a planned change to the model. Comparisons may change as providers update models, prices, or evaluation results.
Key Questions
What did Muse Ember score on Artificial Analysis’ Intelligence Index?
Muse Ember (max) scored 36, matching Claude Haiku 5.5 (max) in the results accessed November 10, 2026. The xhigh setting scored 34.
How much did each benchmark task cost?
Artificial Analysis reported an average of $0.14 per index task for Muse Ember (max), compared with $0.41 for Haiku 5.5 (max). The figure reflects the benchmark’s task usage and is not necessarily the cost of a different workload.
Why does the factual-knowledge result come with a caveat?
On AA-Omniscience, Muse Ember answered 21.1% correctly and declined 64% of questions. Artificial Analysis reported a 17.4% hallucination rate, but the frequent declines mean it may leave more questions unanswered.
Is Muse Ember better than Claude Haiku 5.5?
The comparison is mixed. They tied at 36 on the Intelligence Index, while Muse Ember had a lower reported index-task cost and stronger results on some cited tests. Haiku scored higher on AA-Omniscience and FrontierCode 2.0. Which is more suitable depends on the task and how refusals are handled.
What should a company check before using Muse Ember?
Run a comparison on representative work, tracking quality, refusals, speed, and total cost. For tasks that need factual recall, test the model with relevant source material and decide how unanswered requests will be routed.
Source: ThorstenMeyerAI.com
