TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meta released Muse Ember on November 9, 2026. Artificial Analysis scored its max-effort version at 36 on its Intelligence Index and put the average cost of an index task at $0.14, compared with $0.41 for Claude Haiku 5.5 at the same score. The evaluation also found low factual accuracy, frequent refusals and a lower hallucination rate than several compared models.
Meta released Muse Ember on November 9, and an Artificial Analysis evaluation gave its max-effort version an Intelligence Index score of 36 at an average cost of $0.14 per index task. Claude Haiku 5.5 scored the same, at $0.41 per task in the comparison, while the evaluation also found that Ember answered few questions correctly on a factual knowledge test and declined many of them.
Artificial Analysis says the Intelligence Index combines nine evaluations. In its model data accessed November 10, Muse Ember (max) scored 36; the xhigh setting scored 34 and cost $0.12 per index task. The task cost is an average API cost for one task in the index, not a general price for every user request. Artificial Analysis identifies Ember (max) as the least expensive model scoring 36 or higher in its index.
The index comparison shows a gap between per-task cost and listed token prices. Muse Ember and Claude Haiku 5.5 both list input at $0.10 and output at $0.50 per million tokens, while Ember’s cached input price is $0.004 per million tokens, 60% below Haiku’s, according to the source. Artificial Analysis also reports that Ember used roughly one-third as many output tokens as Haiku to complete the index. It measured 440 output tokens per second and 3.2 minutes per index task through Meta’s API; the model has a stated one-million-token context window.
Results varied by task. Ember scored 68.1% on AutomationBench-AA, level with Meta’s more expensive Muse Wonder in that test, and ahead of Haiku’s 35.4%. It scored 40.9% on Terminal-Bench 4.0, compared with Haiku’s 32.8%, but scored 27.2% on FrontierCode 2.0, below Haiku’s 34.9%. On GDP.pdf, Ember scored 22.8% versus Haiku’s 20.8%; on the AA-Briefcase knowledge-work test, Ember’s 1,521 Elo trailed Haiku’s 1,574.
Muse Ember matches the score at a lower task cost
Meta’s Muse Ember scored level with Claude Haiku 5.5 on Artificial Analysis’s Intelligence Index. Its factual accuracy and frequent refusals make real workload testing essential.
Similar index score, lower task cost
Both models scored 36; benchmark task costs do not predict spending on every business workload.
- $0.14 average per index task
- 17.4% hallucination rate
- $0.41 average per index task
- 34.2% hallucination rate
Results vary by task
Ember led Haiku on some tests and trailed on others.
| Evaluation | Muse Ember | Haiku 5.5 |
|---|---|---|
| AutomationBench-AA | 68.1% | 35.4% |
| Terminal-Bench 4.0 | 40.9% | 32.8% |
| FrontierCode 2.0 | 27.2% | 34.9% |
| GDP.pdf | 22.8% | 20.8% |
Factual answers are a caveat
In AA-Omniscience, Ember answered few questions correctly and declined many.
What businesses should check
Benchmark results are a starting point, not proof of performance in a company’s own work.
Test prompts and workflows drawn from the work the model would handle.
Check factual accuracy and how often refusals interrupt the task.
Include prompts, context, output length, caching and usage patterns.
Lower Costs Come With Trade-offs
The result gives businesses a new price and performance point to evaluate for repeated model tasks. If a workload resembles the index and can use Ember’s strengths, its $0.14 average task cost could make it less expensive to run than Haiku 5.5 at $0.41 under the reported test conditions. Artificial Analysis attributes much of the cost difference to cheaper cached input and lower output-token use. Those figures describe the benchmark, however; they do not establish savings for a particular company or application.
The model’s behavior on factual questions also matters to deployment choices. In AA-Omniscience, Muse Ember answered 21.1% correctly and declined 64% of questions. Artificial Analysis measured a 17.4% hallucination rate, compared with 34.2% for Haiku 5.5 and 79.3% for GPT-6 Luna in the same data. A system that declines uncertain questions may suit some workflows, but the low correct-answer rate limits its value for tasks that require broad factual recall. Companies would need to assess both the quality of answers and how often a refusal interrupts the work.
Products Worth Considering
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Where Ember Fits in Meta’s Line
Muse Ember is the efficient tier of Meta’s Muse model line. It is proprietary and available through Meta’s API and the Muse Code coding agent. The model accepts text and images and produces text. Artificial Analysis tested it at max and xhigh reasoning effort, with the max setting scoring two index points higher while costing two cents more per index task in the reported data.
Artificial Analysis’s comparison also places Ember well below higher-scoring models in both score and reported task cost. Meta’s Muse Wonder scored 44 at $6.95 per index task; Anthropic’s Claude Fable 5.2 scored 65 at $26.54 at its highest effort level, according to the source data. These figures come from the same publisher’s benchmark data and do not, by themselves, show how the models compare on a specific company’s work. The source describes Ember’s score as matching Haiku’s while emphasizing that results differ across individual tests.
Benchmark Results Leave Questions
The source material does not report independent replications of these results or establish how well the benchmark predicts performance in business deployments. Index costs are averages for its tasks; actual API spending depends on a workload’s prompts, context, output length, caching and usage patterns. The reported token prices and speed also do not show what customers will experience under different conditions.
It is also unclear how often Ember’s refusals would be useful or disruptive in specific applications. The AA-Omniscience findings cover one factual-question test and do not establish the model’s error rate across every subject or task. The comparison data supplied here does not include a detailed methodology for each score, confidence ranges, or results from a business customer using Ember in production.
Test Against Real Workloads
The next practical step for organizations considering Ember is to run it against their own representative tasks. That evaluation can compare answer quality, refusal frequency, speed and total API cost with the alternatives already in use. Workflows involving routine extraction, tagging or form completion may merit testing because they can involve many repeated steps, but the benchmark alone cannot confirm that Ember will perform well on those jobs.
Further Artificial Analysis data, changes to Meta’s model or pricing, and customer evaluations could clarify where Ember fits. The source does not give a date for another benchmark update or a planned product milestone. For now, the available evidence is a published benchmark comparison and Meta’s API availability; results in particular applications remain to be established.
Key Questions
What score did Muse Ember receive?
Artificial Analysis scored Muse Ember at max reasoning effort 36 on its Intelligence Index, which combines nine evaluations. Its xhigh setting scored 34 in the data accessed November 10, 2026.How much does an index task cost?
Artificial Analysis reports an average cost of $0.14 per index task for Muse Ember (max). The figure is specific to its benchmark tasks and does not predict the cost of every API request or business workload.How did Ember compare with Claude Haiku 5.5?
Both scored 36 on the reported index. Artificial Analysis lists the average task cost as $0.14 for Ember and $0.41 for Haiku 5.5, while individual tests showed mixed results across workflows, documents and coding.Does Muse Ember often refuse questions?
In Artificial Analysis’s AA-Omniscience test, Ember declined 64% of questions and answered 21.1% correctly. The test also reported a 17.4% hallucination rate; these results apply to that evaluation.Source: Thorsten Meyer AI
