🔍 Read the full analysis: Meta’s Muse Ember AI Costs $0.14 Per Task, But Has A Notable Catch on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Artificial Analysis says Meta’s Muse Ember (max) scored 36 on its Intelligence Index, matching Claude Haiku 5.5, at an average cost of $0.14 per index task versus $0.41. Its lower factual-question accuracy comes with frequent refusals and a lower reported hallucination rate; results may not predict performance on a buyer’s own work.
Meta released Muse Ember on November 9, and independent evaluator Artificial Analysis reports that its max-effort version scored 36 on the Intelligence Index—the same as Claude Haiku 5.5—at an average of $0.14 per index task, compared with $0.41 for Haiku. The price difference comes with a trade-off: Muse Ember performed poorly on a factual-knowledge test but declined to answer many questions, a pattern associated in the evaluation with fewer hallucinations.
Artificial Analysis describes Muse Ember as the efficient tier in Meta’s Muse model line. The proprietary model accepts text and images as input and produces text. Meta offers it through its API and in its Muse Code coding agent. The benchmark figures cited here are from Artificial Analysis model pages accessed on November 10, 2026, and reflect its tests rather than results on customer workloads.
At max reasoning effort, Muse Ember scored 36 and cost an average $0.14 per task across the Intelligence Index. At xhigh effort, it scored 34 and cost $0.12 per task. Artificial Analysis says the index covers nine evaluations; its task-cost figure is the average API cost of completing one index task. The benchmark also recorded 440 output tokens per second and 3.2 minutes per index task for the max version.
The reported token prices are $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. Artificial Analysis attributes Muse Ember’s lower benchmark task cost partly to cached input priced at $0.004 and partly to its use of roughly one-third as many output tokens as Haiku 5.5 during the index. Those are evaluator-reported results, not a guarantee that every deployment will cost the same.
Meta’s Muse Ember: lower benchmark cost with a factuality trade-off
Artificial Analysis reports that Muse Ember (max) matched Claude Haiku 5.5’s Intelligence Index score at a lower average cost per benchmark task. Frequent refusals and weaker factual accuracy are the notable catch.
Where the cost advantage comes from
These are evaluator-reported benchmark costs, not guaranteed prices for a company’s workload.
Same index score
Muse Ember (max) and Claude Haiku 5.5 (max) both scored 36 on the Intelligence Index.
Lower task cost
Artificial Analysis attributes the lower average partly to cached input and fewer output tokens than Haiku 5.5.
API and Muse Code
Meta offers Muse Ember through its API and Muse Code. It accepts text and images and produces text.
The factual-answer trade-off
On AA-Omniscience, frequent refusals accompanied lower factual accuracy and a lower reported hallucination rate.
| AA-Omniscience result | Muse Ember | Claude Haiku 5.5 |
|---|---|---|
| Correct answers | 21.1% | 34.7% |
| Questions declined | 64% | Not specified |
| Reported hallucination rate | 17.4% | 34.2% |
Performance depends on the task
The index score does not capture every capability or weakness.
- 68.1% on AutomationBench-AA, level with Muse Wonder.
- 40.9% on Terminal-Bench 4.0, above Haiku’s 32.8%.
- 22.8% on the GDP.pdf test, above Haiku’s 20.8%.
- 27.2% on FrontierCode 2.0, below Haiku’s 34.9%.
- Lower factual accuracy and frequent refusals on AA-Omniscience.
- Results may differ on customer prompts, tools, and workloads.
What businesses should test
The evaluation does not establish savings or reliability on a particular deployment.
Test representative prompts, documents, and workflow steps.
Measure accuracy with supplied sources and decide how unanswered requests are handled.
Include prompt length, cached input, output volume, reasoning settings, and latency.
Lower Costs, Uneven Capabilities
The result may interest organizations that run high-volume, repeatable AI tasks, such as tagging, extraction, form filling and routine workflow steps. On Artificial Analysis’ stated benchmark basis, Muse Ember matched Haiku 5.5’s index score for about one-third of the task cost. A lower average benchmark cost could matter at scale, but it does not establish savings for a particular company’s prompts, tools, or usage patterns.
The model’s factual-question behavior points to a different deployment consideration. On Artificial Analysis’ AA-Omniscience test, Muse 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. That combination could suit tasks where a system can supply trusted source material and route unanswered questions elsewhere. It may be a poor fit when users expect reliable answers from the model’s stored knowledge alone.
Performance also varies by task. Artificial Analysis reports a 68.1% score on AutomationBench-AA, level with Meta’s more expensive Muse Wonder, and stronger results than Haiku on Terminal-Bench 4.0. But Muse Ember trailed Haiku on FrontierCode 2.0 and on the AA-Briefcase knowledge-work test. The index score alone does not capture those differences, so teams should evaluate the specific work they plan to automate.
Popular right now
Breville Barista Express Espresso Machine with Built-In Grinder, Black Truffle
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
De'Longhi Classic Signature Espresso Machine with Adjustable Milk Frother
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares
Artificial Analysis’ comparison puts Muse Ember (max) and Claude Haiku 5.5 (max) at the same Intelligence Index score of 36, with reported task costs of $0.14 and $0.41 respectively. The same source lists GPT-6 Luna at score 28 and $0.15 per task, and Meta’s Muse Wonder at score 44 and $6.95. These figures compare performance and API cost under the evaluator’s index methodology; they are not universal prices for all applications.
The reported task-specific results are mixed rather than uniformly favorable to Meta’s model. Muse Ember scored 22.8% on the GDP.pdf document test against Haiku’s 20.8%, and 40.9% on Terminal-Bench 4.0 against 32.8%. On FrontierCode 2.0, it scored 27.2%, below Haiku’s 34.9%. Artificial Analysis also reports a score of 46 for Muse Ember in Meta’s Muse Code on its Coding Agent Index, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex. The tool and model combinations differ, so those scores should be read as benchmark results, not a controlled comparison of every coding environment.
The factual test illustrates the model’s central limitation. Artificial Analysis says Muse Ember’s 21.1% accuracy on AA-Omniscience was the lowest among similarly intelligent models in its comparison. It also reports a 64% refusal rate and 17.4% hallucination rate. In the same test, Haiku answered 34.7% correctly and had a 34.2% hallucination rate; GPT-6 Luna’s reported hallucination rate was 79.3%. The evaluator’s figures describe performance on that test, not the likelihood of errors across all uses.
Limits of the Price Comparison
Artificial Analysis’ figures do not establish how Muse Ember will perform or what it will cost on any particular customer workload. The article’s $0.14 figure is an average cost per task in the evaluator’s index; actual API bills can vary with prompt length, cached-input use, output volume, reasoning settings and task design. The source does not provide independent customer deployment results.
It is also unclear how the model’s refusal behavior will affect day-to-day use across different prompts and domains. The AA-Omniscience result is one test of difficult factual questions, not a broad measure of all knowledge work. The supplied material does not report longer-term reliability, customer adoption, or how Meta’s model compares under identical tools and operating conditions across every cited benchmark.
Testing Muse Ember in Practice
Businesses considering Muse Ember will need to test it against their own tasks, including answer accuracy, refusal handling, latency and total API cost. For workflows that supply documents or retrieved records, evaluations can measure whether the model uses those materials accurately and when it declines. Teams can also decide whether unanswered requests should go to a human reviewer or another model.
Meta’s release makes Muse Ember available through its API and Muse Code, according to the source material. No additional release date, pricing change, or follow-up benchmark is specified there. Further customer results and independent evaluations would help show whether the reported cost advantage carries over beyond Artificial Analysis’ index.
Key Questions
What did Artificial Analysis report about Muse Ember’s price?
It reported an average $0.14 API cost per Intelligence Index task for Muse Ember at max effort. This is a benchmark figure, not a fixed cost for every real-world task.
How did Muse Ember compare with Claude Haiku 5.5?
Both scored 36 on the Intelligence Index in the reported max-effort comparison. Artificial Analysis listed task costs of $0.14 for Muse Ember and $0.41 for Haiku 5.5, while results differed across individual tests.
What is the catch with Muse Ember?
On the AA-Omniscience factual-question test, it answered 21.1% correctly and declined 64% of questions. Artificial Analysis reported a 17.4% hallucination rate on that test, so its lower reported hallucination rate came alongside weaker factual accuracy and frequent refusals.
Where is Muse Ember available?
The source says Meta offers the proprietary model through its API and Muse Code. It accepts text and images as input and generates text.
Does the benchmark prove businesses will save money?
No. The $0.14 figure reflects Artificial Analysis’ index tasks. Actual costs and results depend on a company’s workload, prompt and output sizes, use of cached input, and model settings; organizations would need to test those conditions directly.
Source: ThorstenMeyerAI.com
