🔍 Read the full analysis: The Muse Ember Catch: 36 Score, $0.14 Tasks, And Frequent Doubt on ThorstenMeyerAI.com
Get the latest gadgets 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 measured its top tested setting at 36 on its Intelligence Index, matching Claude Haiku 5.5 while costing $0.14 per index task versus $0.41; the benchmark also found lower factual accuracy but a lower hallucination rate. Results are from Artificial Analysis’s tests, not a guarantee of performance or cost on other workloads.
Meta released Muse Ember on November 9, and Artificial Analysis reports that its max setting scored 36 on the Intelligence Index, matching Claude Haiku 5.5 at a lower benchmark cost. The evaluator puts Muse Ember at $0.14 per index task, compared with $0.41 for Haiku 5.5, while its factual-knowledge test results show a trade-off: it answered fewer questions correctly but also hallucinated less often.
Artificial Analysis says Muse Ember is the efficient tier of Meta’s Muse model line, available through Meta’s API and in the Muse Code coding agent. It accepts text and images as input and produces text. The company tested two reasoning settings: max, which scored 36, and xhigh, which scored 34. At xhigh, the reported cost per index task was $0.12, with an average task time of 2.8 minutes; the max setting cost $0.14 and took 3.2 minutes.
The per-task cost is not the same as the listed token price. Artificial Analysis lists Muse Ember at $0.10 per million input tokens and $0.50 per million output tokens, matching the listed prices for Haiku 5.5 and GPT-6 Luna. Its cached-input price is $0.004 per million tokens, which the evaluator describes as 60% below Haiku’s. Artificial Analysis attributes the lower benchmark task cost to cheaper cached input and Muse Ember using roughly a third as many output tokens as Haiku 5.5 across the index.
Results vary by test. Artificial Analysis reports 68.1% on AutomationBench-AA, equal to Meta’s more expensive Muse Wonder in that evaluation and above Haiku 5.5’s 35.4%. Muse Ember scored 40.9% on Terminal-Bench 4.0 versus Haiku’s 32.8%, but scored 27.2% on FrontierCode 2.0, below Haiku’s 34.9%. These are benchmark results reported by Artificial Analysis, not independent evidence of performance on every customer’s tasks.
Muse Ember: Score 36 at a lower benchmark cost
Meta released Muse Ember on November 9, 2026. Artificial Analysis reports a score matching Claude Haiku 5.5, with lower cost per index task but weaker factual accuracy and more frequent refusals.
Where the score comes from
Artificial Analysis combines results across nine evaluations; individual tests show different strengths.
Same index score
Muse Ember and Claude Haiku 5.5 both scored 36 at max effort.
AutomationBench-AA
Muse Ember scored 68.1%, compared with Haiku 5.5 at 35.4%.
FrontierCode 2.0
Muse Ember scored 27.2%, below Haiku 5.5 at 34.9%.
The factual-answer trade-off
On AA-Omniscience, the evaluator found lower accuracy alongside fewer hallucinations than for Haiku 5.5.
| Measure | Muse Ember | Claude Haiku 5.5 |
|---|---|---|
| Accuracy | 21.1% | 34.7% |
| Hallucination rate | 17.4% | 34.2% |
| Refusal rate | 64% | Not provided |
Check fit before switching
Benchmark costs and results may not carry over to a company’s prompts, data, or workflow.
Compare Muse Ember with alternatives using the same prompts and success criteria.
Include retries, fallback model calls, and human review alongside latency and accuracy.
Test whether supplied documents or retrieved information improve results.
A Lower-Cost Option for Repeated Tasks
The result could matter to organizations that run high volumes of routine model calls, such as document extraction, tagging, form filling, or steps within an automated workflow. On Artificial Analysis’s index, Muse Ember’s reported cost per task is about one-third of Haiku 5.5’s. If comparable cost differences hold for a particular workload, shifting suitable tasks could reduce spending; the benchmark alone cannot establish that it will.
The model’s reported tendency to decline questions also affects where it may fit. Artificial Analysis found 21.1% accuracy on AA-Omniscience, alongside a 64% refusal rate and a 17.4% hallucination rate. Haiku 5.5 scored 34.7% accuracy and had a 34.2% hallucination rate in the same comparison. For applications that can supply source material and route unanswered cases to another system, refusing to guess may be useful. For tasks requiring reliable recall without supplied information, the low accuracy is a limitation.
Those findings make workload-specific testing important. A benchmark score compresses results from particular evaluations; it does not show whether a model can follow a company’s rules, use its data correctly, or meet its latency and reliability requirements. Buyers should compare complete costs and error rates on representative tasks, including the cost of retries, human review, and fallback models.
Popular right now
Samsung 43-Inch M70H Series 4K Mini LED Smart TV (2026)
Philips 3300 Series Fully Automatic Espresso Machine with Milk Frother, Grinder, Hot and Iced Coffee Presets, SilentBrew, 15-Bar Pressure (E
Vitamix Touch & Go Advance 32-Ounce Bar Blender, White
As an affiliate, we earn on qualifying purchases.
How the Index Compares Models
Artificial Analysis describes its Intelligence Index as a measure across nine evaluations. Its published comparison puts Muse Ember (max) and Claude Haiku 5.5 (max) at 36. The same source lists Muse Wonder (max) at 44 for $6.95 per index task and Claude Fable 5.2 at 65 for $26.54 at its highest effort setting. These figures illustrate a cost-and-score comparison within the evaluator’s setup; they do not establish that one model is best for every use.
The benchmark also reports a one-million-token context window for Muse Ember and an output speed of 440 tokens per second at max. Artificial Analysis’s reported speed and task duration are measurements from its tests, and actual response time can vary with prompts, system load, and implementation. Meta’s model is proprietary, so the source material does not describe its training data or provide a way for readers to reproduce the model itself.
Benchmark Results Have Limits
The reported results come from Artificial Analysis’s evaluations; the supplied material does not include a separate independent replication or Meta’s own detailed methodology for the claims. It is not clear how Muse Ember will perform across different prompts, languages, industries, or production systems, or whether its benchmark cost advantage will carry over to real workloads with different context sizes and caching patterns.
The comparison also does not resolve the practical effect of the model’s high refusal rate. A refusal can prevent an unsupported answer, but frequent declines may interrupt a workflow or require another model or human reviewer. The source material does not quantify how often users would need such a fallback, or the additional costs that would create.
Test Costs on Real Workloads
The next useful step for potential users is to evaluate Muse Ember against their own tasks, using the same prompts and success criteria as competing models. Tests should track accuracy, refusal rates, latency, and total cost, including any fallback calls or human review. Organizations using it for factual work should also check whether providing documents or retrieved information improves results.
Meta’s API and Muse Code are the access points identified in the source material. Further public information about availability, operational performance, and independent testing may clarify how well the benchmark findings translate beyond Artificial Analysis’s index. Until then, the score and cost figures should be treated as evaluator-reported measurements rather than guarantees.
Key Questions
What did Meta release?
Meta released Muse Ember on November 9, 2026. The model is available through Meta’s API and in its Muse Code coding agent, according to the source material.
How did Muse Ember compare with Claude Haiku 5.5?
Artificial Analysis gave both models a score of 36 on its Intelligence Index at max effort. It reported a cost of $0.14 per index task for Muse Ember and $0.41 for Haiku 5.5.
Does the lower benchmark cost mean it will be cheaper for every use?
No. The $0.14 figure is the evaluator’s average cost for one task in its index. Costs for other workloads may differ based on input and output length, caching, and how the model is used.
What is Muse Ember’s main reported limitation?
Artificial Analysis found 21.1% accuracy on AA-Omniscience, while the model declined 64% of questions. That test suggests limited performance on difficult factual questions, even as the evaluator measured a lower hallucination rate than for the listed comparison models.
Source: ThorstenMeyerAI.com
