🔍 Read the full analysis: Meta Muse Ember Gets A 36 Score, But Its Answers Can Be Uncertain on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Meta released Muse Ember on November 9, 2026. Artificial Analysis gave its max-effort version an Intelligence Index score of 36, matching Claude Haiku 5.5, while estimating an index-task cost of $0.14 versus $0.41 for Haiku. The benchmark also found lower factual accuracy for Muse Ember, alongside a higher refusal rate and a lower hallucination rate.
Meta released Muse Ember on November 9, and an Artificial Analysis evaluation gave its max-effort version an Intelligence Index score of 36—the same score as Claude Haiku 5.5. Artificial Analysis estimated that an index task cost $0.14 with Muse Ember, compared with $0.41 for Haiku, while also finding that Ember answered fewer factual questions correctly and more often declined to answer.
The figures come from Artificial Analysis’s nine-evaluation Intelligence Index, not from a test of every possible use. Its model pages list two Muse Ember reasoning settings: max, which scored 36 at an estimated $0.14 per index task, and xhigh, which scored 34 at $0.12. The source describes task cost as the average API cost of one task in the index.
The listed token prices are $0.10 per million input tokens and $0.50 per million output tokens. Artificial Analysis says Ember’s cached input costs $0.004, 60% less than Haiku 5.5’s cached input, and that Ember used roughly a third as many output tokens as Haiku to complete the index. The source identifies lower cache costs and lower output-token use as reasons for the difference in estimated task cost.
Performance varied by evaluation. Ember scored 68.1% on AutomationBench-AA, matching Meta’s more expensive Muse Wonder on that test 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%. On the AA-Briefcase knowledge-work test, its Elo rating was 1,521, below Haiku’s 1,574.
Meta Muse Ember Gets a 36 Score, but Answers Can Be Uncertain
Released November 9, 2026, Muse Ember matched Claude Haiku 5.5 on Artificial Analysis’s Intelligence Index at a lower estimated benchmark-task cost. Its factual-answer results show an important trade-off for businesses.
Similar score, lower benchmark cost
Artificial Analysis estimated the cost of one index task; it is not a price guarantee for a business workload.
| Model | Index score | Estimated task cost |
|---|---|---|
| Muse Ember, max | 36 | $0.14 |
| Claude Haiku 5.5, max | 36 | $0.41 |
Factual answers need careful handling
On AA-Omniscience, Ember declined more often but had a lower reported hallucination rate than Haiku 5.5.
Performance varies by task
Ember led Haiku on some coding and automation evaluations, but trailed on others.
- AutomationBench-AA: 68.1%, versus Haiku’s 35.4%.
- Terminal-Bench 4.0: 40.9%, versus Haiku’s 32.8%.
- FrontierCode 2.0: 27.2%, versus Haiku’s 34.9%.
- AA-Briefcase Elo: 1,521, below Haiku’s 1,574.
Test your own workflow
Benchmark results are a starting point, not proof of performance on company data or processes.
Use representative tasks and include cases with supplied facts and questions requiring outside knowledge.
Measure correct answers, refusals, hallucinations, response time and full API cost.
Check whether human review or a fallback model is needed, and account for the added process.
Lower Cost Comes With Trade-Offs
The result gives organizations a potential lower-cost option for repetitive, high-volume tasks, such as extraction, tagging, form filling and steps in automated workflows. But a benchmark price is not a quote for a business’s actual workload: costs can depend on input length, output length, caching and how many model steps a task requires. Artificial Analysis’s comparison is a reason to test the model, not proof that it will cost one-third as much in every deployment.
Ember’s factual-question results point to a different kind of trade-off. On AA-Omniscience, it answered 21.1% of questions correctly and declined 64%; its reported hallucination rate was 17.4%. Haiku 5.5 answered 34.7% correctly and had a 34.2% hallucination rate in the same comparison. Those results suggest Ember may suit tasks where the needed information is supplied in a document or retrieved from a trusted source, but could be a poor fit when users expect it to answer obscure questions from stored knowledge.
For companies considering the model, the practical question is whether its refusals are manageable. A system can route unanswered questions to a person or a second model, but that adds process and may add cost. Teams would need to measure both answer quality and how often fallback handling is triggered on their own tasks.
Products Worth Considering
Dell UltraSharp U2424H 24-Inch FHD Monitor
Samsung 32 Odyssey QD-OLED G8 Gaming Monitor – 4K, 240Hz
Samsung 32-inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares Models
Muse Ember is the efficient tier in Meta’s Muse model line. According to the source material, it is a proprietary model available through Meta’s API and in Muse Code, Meta’s coding agent. It accepts text and images and generates text, and has a reported context window of one million tokens. Artificial Analysis tested it at max and xhigh reasoning effort; the scores and cost figures in the headline comparison refer to max.
In the same Artificial Analysis comparison, Muse Wonder scored 44 at an estimated $6.95 per index task. The source also reports Claude Fable 5.2 at 65 and $26.54 per task at its highest effort level. These figures illustrate how benchmark scores and estimated task costs differ across models, but they do not establish which model is best for a specific application.
The reported speed for Muse Ember at max was 440 output tokens per second, with an average 3.2 minutes per index task on Meta’s API. At xhigh, the reported speed was 423 tokens per second and the task took 2.8 minutes. The source does not provide enough information here to treat those timing figures as a prediction for every user’s latency.
Questions About Real-World Performance
The benchmark does not show how Muse Ember performs on any particular company’s data, prompts or workflow. Real-world accuracy, total operating cost and latency remain uncertain without tests in those settings. The source advises measuring performance on a team’s own tasks rather than assuming the index reflects its workload.
The figures also do not explain in detail how the benchmark’s estimates would translate into a production system with retries, human review or fallback models. Nor does the supplied material establish how often refusals would interrupt specific business processes. The underlying evaluations and product terms may change; the reported scores and prices are tied to Artificial Analysis data accessed November 9 and 10, 2026.
Testing Ember on Company Workflows
The next useful step for prospective users is to run side-by-side evaluations on their own tasks, tracking correct answers, refusals, hallucinations, response time and full API cost. Tests should include cases where the model receives all necessary facts, as well as questions that require outside knowledge, to show where retrieval or human review is needed.
Meta’s API and Muse Code are the stated access routes in the source material. The supplied information does not give a broader rollout schedule, product availability terms or a date for further benchmark updates. Until those details and independent task-specific results are available, the score and cost comparison should be read as a benchmark snapshot rather than a guarantee of performance.
Key Questions
What score did Muse Ember receive?
Artificial Analysis gave Muse Ember at max reasoning effort a score of 36 on its Intelligence Index. Its xhigh version scored 34.
How does its estimated task cost compare with Claude Haiku 5.5?
Artificial Analysis estimated $0.14 per index task for Muse Ember (max) and $0.41 for Claude Haiku 5.5 (max). These are benchmark task estimates, not guaranteed costs for other workloads.
Does Muse Ember answer factual questions reliably?
In Artificial Analysis’s AA-Omniscience test, Ember answered 21.1% correctly and declined 64% of questions. The result suggests caution when a task depends on the model’s stored factual knowledge.
What does the benchmark report about hallucinations?
Artificial Analysis reported a 17.4% hallucination rate for Muse Ember in the AA-Omniscience comparison. That was below the 34.2% reported for Haiku 5.5 in the same comparison; benchmark results do not guarantee performance in other settings.
Where is Muse Ember available?
The source material says the proprietary model is available through Meta’s API and Muse Code. It does not provide more detailed availability or rollout terms.
Source: ThorstenMeyerAI.com
