🔍 Read the full analysis: Muse Ember’s AI Results: 36 Score, $0.14 A Task, Frequent Doubt 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 scores its max-effort version at 36 on a nine-evaluation index and estimates an average cost of $0.14 per index task, compared with $0.41 for Claude Haiku 5.5 at the same score. The benchmark also found low factual accuracy but a high refusal rate and a lower hallucination rate than Haiku 5.5.
Meta released Muse Ember on November 9, and independent benchmark provider Artificial Analysis reports that its max-effort version scored 36 on the Intelligence Index, matching Claude Haiku 5.5 while costing an estimated $0.14 per index task versus $0.41. The comparison points to a potentially lower-cost option for some AI workloads, but the benchmark also found that Ember answered few questions correctly on a factual-knowledge test and often declined to answer.
Artificial Analysis says its Intelligence Index combines results from nine evaluations. In the provider’s model-page data accessed November 10, Muse Ember (max) scored 36 and cost an average of $0.14 to complete one task in the index. Claude Haiku 5.5 (max) also scored 36, at an estimated $0.41 per task. These are benchmark-based estimates, not guarantees of the price or performance users will see on their own workloads.
The per-task gap does not come from lower listed token rates, according to the source: Muse Ember and Haiku 5.5 are both listed at $0.10 per million input tokens and $0.50 per million output tokens. Artificial Analysis attributes Ember’s lower benchmark cost partly to cached input priced at $0.004 and partly to its use of roughly a third as many output tokens as Haiku 5.5 in the index. The source reports 440 output tokens per second and an average 3.2 minutes per task for Ember (max) on Meta’s API.
Results varied by task. Ember scored 68.1% on AutomationBench-AA, matching Meta’s more expensive Muse Wonder on that test and exceeding Haiku 5.5’s 35.4%. It was ahead of Haiku on Terminal-Bench 4.0, but behind on FrontierCode 2.0 and the AA-Briefcase knowledge-work test. Artificial Analysis also reports an Ember score of 46 on its Coding Agent Index when used with Meta’s Muse Code tool; that result compares different model-and-tool setups, so it should not be read as a direct measure of standalone model performance.
Muse Ember matches the score at a lower benchmark cost
Artificial Analysis found Muse Ember (max) tied Claude Haiku 5.5 on its Intelligence Index, but reported weak factual-test accuracy and frequent refusals.
Same score, different cost
The benchmark estimates are tied to Artificial Analysis’s index tasks, not guaranteed prices for customer workloads.
- Score: 36
- Estimated task cost: $0.14
- Lower cost partly reflects cached input and fewer output tokens in the index
- Score: 36
- Estimated task cost: $0.41
- Same listed input and output token rates as Ember
Factual answers remain a concern
On AA-Omniscience, Ember often declined rather than answering—and still answered few questions correctly.
Performance varies by task
Ember led Haiku 5.5 on some tests but trailed it on others, so an overall score is not a task-by-task guarantee.
AutomationBench-AA
Ember scored 68.1%, above Haiku 5.5’s 35.4% on this test.
Other task tests
Ember was ahead on Terminal-Bench 4.0, but behind on FrontierCode 2.0 and AA-Briefcase.
Coding-agent result
Its Coding Agent Index score of 46 used Muse Code, so it is not a standalone model comparison.
Test before routing work
A lower benchmark estimate may suit some workloads, but real costs and results depend on prompts, context, caching, refusals, and fallback handling.
Evaluate Ember on your own prompts and data, especially knowledge-heavy requests.
Measure accuracy, refusals, latency, token use, and human or system fallback costs.
Decide what happens when Ember declines or cannot complete a request.
Lower Cost, Uneven Performance
The report’s practical point is not that Muse Ember outperforms more capable models across the board. It is that per-task cost can differ sharply even when two models receive the same overall index score. For companies running large numbers of routine steps, such as document extraction, tagging, form filling or workflow actions, a lower cost on comparable tasks could affect deployment choices. The index estimate alone cannot establish savings for a particular business: actual costs depend on prompts, context, caching, output length and how a system handles errors.
The accuracy results also make task selection important. On Artificial Analysis’s AA-Omniscience test, Ember answered 21.1% correctly and declined 64% of questions. Its reported hallucination rate was 17.4%, compared with 34.2% for Haiku 5.5 in the same source data. A high refusal rate may be useful where unsupported answers are costly, but it can also mean more unanswered requests or extra work for a fallback system. The figures describe performance on that test, not a general guarantee that Ember will avoid false claims in production.
Products Worth Considering
Samsung Galaxy Tab A11+ 8GB RAM, 256GB Storage, Gray
Amazon Kindle Paperwhite (2024, 16 GB) – 7-inch Glare-Free Display, Black
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
As an affiliate, we earn on qualifying purchases.
Meta’s New Efficient Model Tier
Muse Ember is described in the source as the efficient tier in Meta’s Muse model line. It is proprietary, takes text and images as input, and generates text. Meta makes it available through its API and in Muse Code, its coding agent. Artificial Analysis tested two reasoning-effort settings: max scored 36 on the Intelligence Index, while xhigh scored 34 and had an estimated index-task cost of $0.12.
The benchmark comparisons show a substantial gap between lower-cost models and some higher-scoring options, but they are not a single price-versus-quality ranking for every use. In the data cited by Artificial Analysis, Muse Wonder scored 44 at $6.95 per index task, while Claude Fable 5.2 at its highest effort level scored 65 at $26.54. Those figures provide a benchmark reference point; they do not show that the models are interchangeable or that any one is the right choice for a specific task.
Artificial Analysis’s figures were reported on November 9 and 10, 2026, and reflect its own tests and model settings. The source describes the Intelligence Index as based on nine evaluations, while AA-Omniscience separately tests factual questions and allows models to decline. Keeping those measures distinct matters: a model’s overall index score does not by itself indicate how accurately it will answer obscure factual questions.
Benchmark Results Have Limits
The available figures do not establish how Muse Ember will perform across customers’ private datasets, prompts or production systems. Artificial Analysis’s per-task cost is an estimate tied to its index; the source does not provide a breakdown of how that estimate may change with different context lengths, caching patterns or API use. It also does not report a direct cost comparison for every task type in the benchmark.
It remains unclear how frequently users will encounter Ember’s refusals outside AA-Omniscience, how those declines affect end-to-end workflow costs, and whether the reported benchmark results will remain stable as model versions or API conditions change. The source gives no independent evaluation beyond the cited testing, and its coding-agent comparison involves different model-and-tool configurations. Readers should treat the reported scores as a snapshot, not a broad prediction of performance.
Test Before Routing Workloads
The next practical step for organizations considering Muse Ember is to evaluate it on their own representative tasks, tracking accuracy, refusal rates, latency, token use and the cost of any fallback or human review. For knowledge-heavy requests, a system may need to provide relevant documents or retrieved records to the model rather than rely on its stored knowledge. Any deployment should also account for cases where Ember declines and decide which system, if any, handles those requests.
Meta’s API and Muse Code are the availability routes identified in the source. Further comparisons will depend on additional independent testing and updates to the models and benchmarks. The current report does not say when Artificial Analysis will publish a follow-up or whether Meta will change pricing, so both remain unconfirmed.
Key Questions
What did Artificial Analysis report about Muse Ember?
Artificial Analysis reported that Muse Ember (max) scored 36 on its nine-evaluation Intelligence Index and had an estimated cost of $0.14 per index task. The figures were accessed November 10, 2026.
How does Muse Ember’s benchmark cost compare with Claude Haiku 5.5?
Both models scored 36 on the Intelligence Index in the cited comparison. Artificial Analysis estimated $0.14 per task for Muse Ember (max) and $0.41 for Haiku 5.5 (max); these are index-task estimates, not prices for every real-world request.
Does Muse Ember answer factual questions reliably?
On AA-Omniscience, Artificial Analysis reported 21.1% correct answers and a 64% refusal rate for Muse Ember. Those results indicate weak performance on that factual-knowledge test and do not establish accuracy across all uses.
Why was Muse Ember’s reported hallucination rate lower?
Artificial Analysis reported a 17.4% hallucination rate for Ember on AA-Omniscience and said it declined many questions rather than guessing. The source reports 34.2% for Haiku 5.5 on the same test; these figures apply to that evaluation.
Where can users access Muse Ember?
The source says Meta offers Muse Ember through its API and Muse Code, Meta’s coding agent. It does not provide further account, regional availability or access terms.
Source: ThorstenMeyerAI.com
