🔍 Read the full analysis: Muse Ember Scores 36 For 14 Cents Per Task—How Often Does It Admit Uncertainty? on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Meta released Muse Ember on November 9, 2026. Artificial Analysis reports that its max-effort version scored 36 on the Intelligence Index at an average $0.14 per task, matching Claude Haiku 5.5’s score at $0.41. In a separate factual-knowledge test, Ember answered correctly less often but declined more questions and had a lower measured hallucination rate.
Meta released Muse Ember on November 9, and testing by Artificial Analysis found its max-effort version scored 36 on the firm’s Intelligence Index for an average $0.14 per index task. Claude Haiku 5.5 also scored 36, but its reported task cost was $0.41, making Ember a lower-cost option on this particular benchmark while showing weaker results on factual knowledge.
Artificial Analysis says its Intelligence Index combines nine evaluations. In its model data accessed November 10, Muse Ember at max reasoning effort scored 36; its xhigh setting scored 34 at an average $0.12 per index task. The max version produced output at a reported 440 tokens per second and took an average 3.2 minutes per index task on Meta’s API. Those are benchmark measurements, not guarantees of performance or speed on a user’s workload.
The source reports list prices of $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. It attributes Ember’s lower overall index-task cost in part to cached input priced at $0.004 and to using roughly a third as many output tokens as Haiku across the index. Actual API bills will depend on the task and usage pattern.
Results varied by test. Ember scored 68.1% on AutomationBench-AA, level with Meta’s more expensive Muse Wonder according to the supplied data, and beat Haiku 5.5 on Terminal-Bench 4.0, 40.9% to 32.8%. It trailed Haiku on FrontierCode 2.0, 27.2% to 34.9%. The source also reports a Coding Agent Index score of 46 for Ember in Meta’s Muse Code tool; that result compares tool-and-model combinations, rather than measuring the model in isolation.
Muse Ember: matched score, lower benchmark cost
Meta’s Muse Ember matched Claude Haiku 5.5 on Artificial Analysis’ Intelligence Index at a lower average cost per task. Its factual-knowledge results show a different trade-off: more declines and a lower measured hallucination rate.
Benchmark cost, not a universal price
The reported index-task cost is specific to Artificial Analysis’ benchmark setup; real bills depend on the task and usage pattern.
- Index score: 36
- Average index-task cost: $0.14
- Index score: 36
- Average index-task cost: $0.41
Factual answers and uncertainty
On AA-Omniscience, Ember answered correctly less often, declined more questions and had a lower measured hallucination rate than Haiku 5.5.
| AA-Omniscience measure | Muse Ember | Haiku 5.5 |
|---|---|---|
| Correct answers | 21.1% | 34.7% |
| Declined questions | 64% | Not reported here |
| Measured hallucination rate | 17.4% | 34.2% |
Performance depends on the test
Ember’s results varied across coding and automation evaluations, so one index score cannot predict every workload.
AutomationBench-AA
Ember scored 68.1%, level with Muse Wonder in the supplied data.
Terminal-Bench 4.0
Ember scored 40.9%, ahead of Haiku 5.5 at 32.8%.
FrontierCode 2.0
Ember scored 27.2%, below Haiku 5.5 at 34.9%.
Where teams might test Ember
The reported profile may be worth evaluating in workflows that provide source information and can handle unanswered requests.
Try document extraction, tagging, routing or form filling on your own data.
Track accuracy, refusals, latency and total cost against alternatives.
Route unanswered requests to a human reviewer or another model when needed.
Lower Cost Comes With Trade-Offs
The reported price difference could matter to organizations running high-volume, repeated AI tasks, such as document extraction, tagging, routing and form filling. At the measured index-task costs, Ember’s run was about one-third the price of Haiku 5.5’s. That comparison is specific to Artificial Analysis’ benchmark setup; it does not establish that Ember will be cheaper or equally capable on every company’s workload.
The factual-question results point to a separate deployment choice. Artificial Analysis measured Ember at 21.1% accuracy on AA-Omniscience, with declines on 64% of questions and a 17.4% hallucination rate. The source reports Haiku 5.5 at 34.7% accuracy and a 34.2% hallucination rate on that test. A system that often declines may suit workflows with reliable source documents and human or automated fallback, but it may frustrate users who expect an answer from general knowledge. These figures are test-specific and should not be treated as universal rates of correctness or hallucination.
Popular right now
LG 77-Inch OLED evo C6 4K Smart TV
Ninja Foodi Air Fryer 8 Qt DualZone 6-in-1 Grey DZ201
Vitamix Touch & Go Advance 32-Ounce Bar Blender, White
As an affiliate, we earn on qualifying purchases.
How Ember Fits Meta’s Model Line
Muse Ember is described in the source as the efficient tier of Meta’s Muse model line, offered through Meta’s API and the Muse Code coding agent. It accepts text and images as input and generates text. Artificial Analysis tested two reasoning-effort settings, max and xhigh; the headline comparison with Haiku uses the max setting.
The price-and-score comparison has important limits. Artificial Analysis’ task cost is an average for one task in its own index, not a standard unit of work across providers. The source says Muse Wonder scored 44 at $6.95 per index task, while Claude Fable 5.2 scored 65 at $26.54 at its highest effort setting. These figures illustrate differences in the reported benchmark and cost, but do not show what the models would cost or achieve on a particular production workflow.
On AA-Omniscience, the test that permits models to decline difficult factual questions, the supplied results show Ember answering fewer questions correctly than the compared models but declining more often. Accuracy, refusal behavior and hallucination rate measure different things: a lower hallucination rate does not mean a model knows more, and a refusal is not itself proof that the model has identified an unknowable question.
Benchmark Limits and Open Questions
The supplied reporting does not establish how Muse Ember performs across independent real-world deployments, or whether the benchmark cost gap will hold for different prompts, context lengths, caching rates and tool-use patterns. Artificial Analysis’ index is one evaluation; the source itself advises testing models on a user’s own tasks before choosing a high-volume tier.
Details also remain limited on the methods behind the reported refusal and hallucination rates, including how questions were selected and how a refusal was classified. The figures should be read as outcomes on AA-Omniscience, not as a prediction that Ember will refuse 64% of ordinary requests. The source does not provide independent replication, broad customer results, or a full account of availability and service terms beyond access through Meta’s API and Muse Code.
Testing on Real Workloads
For teams considering Ember, the next step is to compare it with alternatives on representative tasks, tracking accuracy, refusal rates, latency and total cost rather than relying on the index score alone. Workflows can also be designed to provide relevant documents or retrieved facts and route unanswered requests to another model or a human reviewer.
Further evidence will depend on additional benchmark results and reporting from users as the model is put into production. The source material gives no date for a new evaluation or product milestone, so whether Ember’s cost and performance profile persists beyond the current tests remains open.
Key Questions
What did Muse Ember score on the Artificial Analysis Intelligence Index?
Its max-effort version scored 36 on the index, which the source says combines nine evaluations. The xhigh version scored 34.
How does its reported task cost compare with Claude Haiku 5.5?
Artificial Analysis reported an average of $0.14 per index task for Muse Ember (max) and $0.41 for Claude Haiku 5.5 (max). This is the cost on that benchmark, not a quote for every use case.
Does Muse Ember have stronger factual knowledge than its rivals?
Not in the supplied AA-Omniscience results. Ember answered 21.1% correctly, below the cited Haiku 5.5 result of 34.7%. It declined 64% of questions and had a reported 17.4% hallucination rate on that test.
What kinds of work might suit Muse Ember?
The reported results suggest it may be worth testing for high-volume workflows that provide their own source information, including extraction, tagging and document-based tasks. Teams should measure performance on their own data and plan how to handle refusals.
Are the benchmark results independently confirmed across real-world use?
The supplied source reports Artificial Analysis benchmark findings, but does not provide independent replication or broad deployment results. Performance and cost on specific workloads remain unconfirmed.
Source: ThorstenMeyerAI.com
