🔍 Read the full analysis: How Muse Ember Scores 36 While Often Admitting It Doesn’t Know on ThorstenMeyerAI.com
Get smart everyday buys 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
Artificial Analysis scored Meta’s Muse Ember (max) at 36 on its Intelligence Index, the same as Claude Haiku 5.5, while reporting an average cost of $0.14 per index task versus $0.41 for Haiku. The benchmark also found lower factual accuracy but a higher rate of refusals and a lower hallucination rate on its AA-Omniscience test. These results describe specific evaluations, not guaranteed performance on other workloads.
Meta’s Muse Ember (max) scored 36 on Artificial Analysis’s Intelligence Index, matching Claude Haiku 5.5 while costing an average of $0.14 per index task, compared with $0.41 for Haiku, according to the benchmark data published November 9-10. The results put a lower-cost option in the same score range, but also show a trade-off: Muse Ember answered fewer questions correctly on a separate factual-knowledge test and declined more often.
Artificial Analysis reports that Muse Ember is the efficient tier of Meta’s Muse model line, released November 9. The proprietary model is available through Meta’s API and in Muse Code, the company’s coding agent. It accepts text and images and generates text. In the index evaluation, the max reasoning setting scored 36; the xhigh setting scored 34 and had a reported average task cost of $0.12.
The benchmark’s model pages list Muse Ember’s input and output prices at $0.10 and $0.50 per million tokens, respectively, the same rates reported for Haiku 5.5 and GPT-6 Luna. Artificial Analysis attributes Ember’s lower average index-task cost relative to Haiku to its cache pricing and lower output-token use: it says cached input costs $0.004 per million tokens, 60% less than Haiku’s, and that Ember used roughly a third as many output tokens as Haiku to complete the index.
Performance varied by test. Artificial Analysis reports 68.1% on AutomationBench-AA, level with its score for the more expensive Muse Wonder, and 40.9% on Terminal-Bench 4.0, ahead of Haiku’s 32.8%. Ember scored 27.2% on FrontierCode 2.0, below Haiku’s 34.9%. In the Artificial Analysis Coding Agent Index, the source reports scores of 46 for Muse Ember in Muse Code, 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex.
How Muse Ember Scores 36 While Often Admitting It Doesn’t Know
Artificial Analysis found Muse Ember (max) matched Claude Haiku 5.5 on its Intelligence Index at a lower average benchmark-task cost. Separate factual-knowledge results show a trade-off in accuracy and refusals.
Same Index Score, Lower Task Cost
The reported costs apply to Artificial Analysis’s benchmark setup, not every prompt or customer workload.
| Model | Index score | Average task cost |
|---|---|---|
| Muse Ember (max) | 36 | $0.14 |
| Claude Haiku 5.5 (max) | 36 | $0.41 |
Factual Test: Accuracy and Refusals
On AA-Omniscience, Ember answered fewer questions correctly but had a lower reported hallucination rate than Haiku.
Performance Varies by Task
Other benchmark results show why a single index score cannot predict performance across workloads.
68.1% on AutomationBench-AA
Level with Muse Wonder’s reported score on this evaluation.
40.9% on Terminal-Bench 4.0
Above Haiku’s reported 32.8% score.
27.2% on FrontierCode 2.0
Below Haiku’s reported 34.9% score.
What Teams Should Test
Use your own workflows to check whether the benchmark trade-offs matter in practice.
Include work based on supplied documents and cases where the model should decline.
Track correctness, unsupported claims, refusals, output volume, latency and total API cost.
For factual tasks, define when to use retrieval, another model or human review.
Lower Cost, With a Knowledge Trade-off
The comparison matters to developers and businesses evaluating models for repeated, high-volume tasks, such as routing, extraction, tagging or filling forms. Artificial Analysis’s $0.14 figure is the average API cost of one task in its index, not a general price for every prompt or a prediction of a customer’s bill. Actual costs will depend on the workload, model settings, input length and use of cached context.
The knowledge test points to a separate deployment consideration. On AA-Omniscience, Artificial Analysis reports that Muse Ember answered 21.1% of questions correctly, declined 64%, and had a 17.4% hallucination rate. Haiku answered 34.7% correctly and had a 34.2% hallucination rate in the same comparison. Those figures suggest Ember may be more likely to withhold an answer on that test, but refusals do not by themselves establish that a model will be safer or more useful in a particular product.
For tasks where the relevant facts are supplied in documents or retrieved from a database, the benchmark results could make Ember worth testing. They do not show that it will outperform alternatives on a company’s own data. Teams would need to test accuracy, refusal handling, latency and total costs on their specific workflows before choosing a model.
Products Worth Considering
Apple Mac mini Desktop Computer with M2 Pro, 16GB Memory, 512GB SSD, Gigabit Ethernet
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera (2026, Silver)
Dell XPS Tower 8960 Desktop, Intel Core i7-13700, NVIDIA GeForce RTX 4060, 16GB DDR5, 512GB SSD
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares Models
The Artificial Analysis Intelligence Index aggregates results from nine evaluations, according to the source. A score of 36 is an index result, not a percentage correct or a direct measure of every capability. The reported $0.14 cost is the average API cost to complete one task in that index under the tested setup; it should not be read as a fixed price for unrelated use.
Artificial Analysis also tested Muse Ember at two reasoning effort levels. Its November 10 model-page data lists the max setting at 36 points, $0.14 per task and 440 output tokens per second, with 3.2 minutes per index task. The xhigh setting scored 34 at $0.12 per task, with 423 output tokens per second and 2.8 minutes per task. The source lists a one-million-token context window for Ember.
On the same comparison, Artificial Analysis lists Muse Wonder at 44 points and $6.95 per index task, and Claude Fable 5.2 at 65 points and $26.54 per task at its highest effort level. These are results and costs under the source’s tested conditions; they do not establish that one model is the best choice for every use case.
Limits of the Reported Results
The figures come from Artificial Analysis’s tests and model pages; the supplied source does not provide independent replication or detailed uncertainty ranges for each result. It is also not clear from the source how performance will compare across different prompts, languages, system instructions, or real-world workloads. Benchmark results can change as models and evaluation methods are updated.
The reported refusal and hallucination rates are specific to AA-Omniscience. The source does not establish why Ember declined particular questions, how refusals were categorized, or whether the same pattern appears in other factual tasks. Nor does the benchmark show the full cost of operating a deployed system, including application infrastructure, retries, human review or fallback models.
Meta’s release information in the source identifies the model’s availability through its API and Muse Code, but does not detail access limits, service availability or future pricing changes. The comparisons should be treated as a snapshot of the cited November 2026 data, not a guarantee of future performance or cost.
Test Results Against Real Workloads
The next practical step for teams considering Muse Ember is to run evaluations on their own tasks, including examples where the model must answer from supplied documents and examples where it should decline. Comparisons should track correctness, unsupported claims, refusal rates, output volume, latency and total API cost against the alternatives already in use.
For systems that depend on factual answers, teams can also set a clear fallback path for refusals, such as retrieval, another model or human review. The source material does not give a further benchmark release date or announce a specific follow-up evaluation, so any changes to the results or product availability remain to be confirmed.
Key Questions
What score did Muse Ember receive?
Artificial Analysis reports a score of 36 for Muse Ember at max reasoning effort on its Intelligence Index. The xhigh setting scored 34. The index combines results from nine evaluations, according to the source.
How did Muse Ember’s benchmark cost compare with Haiku 5.5?
The source reports an average cost of $0.14 per index task for Muse Ember (max) and $0.41 for Claude Haiku 5.5 (max). These are benchmark-task costs under Artificial Analysis’s test setup, not a universal charge for every use.
Does Muse Ember hallucinate less than Haiku 5.5?
On Artificial Analysis’s AA-Omniscience test, Muse Ember’s reported hallucination rate was 17.4%, compared with 34.2% for Haiku 5.5. The result applies to that evaluation and does not establish rates across other tasks.
What is Muse Ember’s main reported weakness?
Its factual performance on AA-Omniscience was low relative to the models cited: it answered 21.1% of questions correctly and declined 64%. The source does not show whether those results carry over to tasks where relevant information is provided in documents or retrieved from a database.
Source: ThorstenMeyerAI.com
