🔍 Read the full analysis: Can You Rely On Muse Ember? Its Score Is 36, But It Often Says “I Don’t Know” on ThorstenMeyerAI.com
TL;DR
Meta released Muse Ember on November 9, 2026. Artificial Analysis gave its maximum-reasoning version an Intelligence Index score of 36 and reported an average cost of $0.14 per index task, compared with $0.41 for Claude Haiku 5.5 at the same score. The evaluation also found low factual accuracy on AA-Omniscience, alongside frequent refusals and a lower hallucination rate than Haiku 5.5.
Meta released Muse Ember on November 9, and independent evaluator Artificial Analysis reports that its maximum-reasoning version scores 36 on the Intelligence Index—the same as Claude Haiku 5.5—at an average cost of $0.14 per index task, versus $0.41 for Haiku. The evaluation also found that Ember often declines to answer difficult factual questions: it answered 21.1% correctly on AA-Omniscience and declined 64% of questions.
Artificial Analysis describes Muse Ember as the efficient tier of Meta’s Muse model line. The proprietary model accepts text and images and produces text, and is available through Meta’s API and the company’s Muse Code coding agent. The reported Intelligence Index score is for Muse Ember at maximum reasoning effort; the xhigh version scored 34 and cost $0.12 per index task in the evaluator’s data.
The cost comparison is based on Artificial Analysis’s index workload, not a general price for every user task. The evaluator says Muse Ember’s input-token list price is $0.10 per million tokens and output price is $0.50 per million, matching Haiku 5.5’s listed rates. It attributes Ember’s lower average index-task cost partly to cached input priced at $0.004 per million tokens—60% below Haiku’s—and to using roughly a third as many output tokens as Haiku to complete the index.
Results vary by task. Artificial Analysis measured Ember at 68.1% on AutomationBench-AA, level with Meta’s more expensive Muse Wonder, while it trailed Haiku on FrontierCode 2.0, scoring 27.2% against 34.9%. Ember scored higher than Haiku on Terminal-Bench 4.0, 40.9% to 32.8%. In the evaluator’s coding-agent index, Ember inside Muse Code scored 46, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex.
Can You Rely on Muse Ember? A score of 36, with frequent refusals
Muse Ember matches Claude Haiku 5.5’s Intelligence Index score at a lower reported benchmark cost. Its factual-question results show a trade-off between fewer hallucinations and frequent refusals.
Same score, lower index cost
The reported task cost applies to Artificial Analysis’s index workload, not every business use.
| Model | Index score | Cost per task |
|---|---|---|
| Muse Ember | 36 | $0.14 |
| Claude Haiku 5.5 | 36 | $0.41 |
| Muse Wonder | 44 | $6.95 |
Factual answers come with caveats
On AA-Omniscience, Ember’s lower hallucination rate appeared alongside many declines and a low correct-answer share.
- 21.1% answered correctly
- 64% declined
- 17.4% hallucination rate
- 34.7% answered correctly
- 34.2% hallucination rate
Performance depends on the task
The benchmark results show strengths in some workflows and weaker results in others.
Level with Muse Wonder
Ember scored 68.1% on AutomationBench-AA.
Trailed Haiku on FrontierCode
Ember scored 27.2%, compared with Haiku’s 34.9%.
Ember scored higher
It reached 40.9% on Terminal-Bench 4.0, versus Haiku’s 32.8%.
Test your own workflow
A benchmark score does not establish performance, completion rates, or costs for a particular organization.
Compare Ember and alternatives on the same inputs and success criteria.
Measure accuracy, refusals, completion time, and total cost.
Count retries, human review, or another model when Ember declines.
Where Lower Cost Changes the Trade-Off
The results position Muse Ember as a potential option for high-volume, repeatable tasks where a business supplies the needed information, such as extraction, tagging, form filling, or steps in an automated workflow. If the evaluator’s cost pattern holds for a company’s own prompts and usage, the lower per-task figure could matter at scale. But the index is a benchmark, not a forecast of a particular organization’s bill, and the cost depends on how much input is cached and how many tokens a task requires.
Its refusal behavior also matters for applications where a confident but incorrect answer carries a cost. Artificial Analysis reports a 17.4% hallucination rate for Ember on AA-Omniscience, compared with 34.2% for Haiku 5.5. That lower rate comes alongside a 64% refusal rate and only 21.1% correct answers on the test. A refusal may be preferable to an unsupported guess, but it can still interrupt a workflow and require a fallback or human review.
The practical choice is not simply “cheaper or more expensive.” Teams would need to compare completion rates, factual reliability, latency, and total cost on their own tasks. Ember’s benchmark profile suggests strengths in some workflows and weaknesses in stored factual knowledge; it does not establish that the model is reliable for every business use.
Popular right now
Shark Rotator Lift-Away Upright Vacuum with DuoClean PowerFins, Wine Purple, LA702
Hisense E6SR Series 55-Inch 4K UHD Hi-QLED Roku Smart TV
Vitamix 5200 Blender with 64 oz. Container, Black
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares Models
The Artificial Analysis Intelligence Index combines nine evaluations into a score. Its published figures compare model versions at specified reasoning-effort levels, and the cost-per-task measure is the average API cost to complete one task in that index. These measures are useful for comparing the tested configurations, but they do not directly predict performance on every prompt, product, or deployment.
Within the comparisons supplied by Artificial Analysis, Claude Haiku 5.5 also scored 36, at $0.41 per index task. Meta’s larger Muse Wonder scored 44 at $6.95 per task. Anthropic’s Claude Fable 5.2 scored 65 at its highest effort level and cost $26.54 per index task in the cited data. Those figures describe the tested models and index workload; they are not a general ranking of quality across all uses.
The factual-knowledge result comes from AA-Omniscience, which tests difficult questions across fields and allows a model to decline. In that evaluation, Haiku 5.5 answered 34.7% correctly, while Gemini 3.8 Flash at high effort answered 50.5% correctly. Ember’s comparatively low hallucination rate should be read alongside its high refusal rate and low correct-answer share, rather than as evidence that it knows more.
What the Scores Do Not Establish
The published results do not show how Muse Ember will perform on any specific company’s data or workflow. Benchmark scores are not guarantees of accuracy, completion rates, or cost in production. The reported $0.14 is an average for one task in Artificial Analysis’s index; actual expense can vary with prompt length, cached input, output volume, and how a task is configured.
The supplied material also does not establish how the model’s refusal rate changes across domains, prompt styles, or versions, or how users will experience it through Meta’s API and Muse Code. It does not provide independent results for every deployment condition. The 17.4% hallucination figure and 64% refusal figure are specific to AA-Omniscience and should not be treated as universal rates.
It remains unclear whether the benchmark cost advantage will persist for real-world workloads, especially where a task needs retries, human checking, or a second model to handle refusals. The available results identify trade-offs, but do not settle whether Ember is the right choice for a particular application.
Testing Ember on Real Workloads
The next useful step for organizations considering Muse Ember is to test it against their own representative tasks. That means measuring accuracy, refusal frequency, completion time, and total cost, including any fallback model or human review required when Ember declines. Results should be compared with alternatives using the same inputs and success criteria.
Meta’s model is available through its API and Muse Code, according to the supplied information. Further independent evaluations and user testing may clarify whether the reported benchmark strengths translate to different deployments. Until then, the evidence supports a narrower conclusion: Muse Ember matches Haiku 5.5’s index score at lower reported cost, while its factual-question results show a substantial willingness to decline.
Key Questions
What score did Muse Ember receive?
Artificial Analysis reports an Intelligence Index score of 36 for Muse Ember at maximum reasoning effort. Its xhigh version scored 34 in the cited results.
Is Muse Ember cheaper than Claude Haiku 5.5?
On Artificial Analysis’s index workload, Muse Ember at max effort averaged $0.14 per task, compared with $0.41 for Haiku 5.5 at max effort. The figures describe that benchmark workload and do not guarantee the same cost for other tasks.
Why does Muse Ember often say it does not know?
Artificial Analysis reports that Ember declined 64% of questions on its AA-Omniscience test, which permits a model to refuse rather than guess. The supplied results do not establish that it will refuse at the same rate in other settings.
Does a lower hallucination rate mean Muse Ember is more knowledgeable?
No. On AA-Omniscience, Ember answered 21.1% correctly and had a reported hallucination rate of 17.4%. Its low hallucination rate occurred alongside frequent refusals; it does not show stronger factual knowledge.
Should businesses rely on Muse Ember?
The reported tests do not establish reliability for a particular business. Organizations would need to evaluate Ember on their own tasks, including its accuracy, refusals, and full operating cost, and decide how to handle questions it cannot answer.
Source: ThorstenMeyerAI.com
