🔍 Read the full analysis: Meta’s Muse Ember Scores 36: Here’s The Catch Behind Its Low Task Cost on ThorstenMeyerAI.com
Get hardware and tech essentials 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 reports that Meta’s Muse Ember scored 36 on its Intelligence Index at max reasoning effort, matching Claude Haiku 5.5 while costing $0.14 per benchmark task versus $0.41. The lower cost comes with a trade-off: tests found weak factual accuracy, though Muse Ember often declined to answer and had a lower measured hallucination rate than several comparators.
Meta released Muse Ember on November 9, and an Artificial Analysis evaluation gave its max-effort version an Intelligence Index score of 36—the same score as Claude Haiku 5.5. The evaluation puts Muse Ember’s average cost at $0.14 per index task, compared with $0.41 for Haiku, but also reports that Ember answers relatively few questions correctly on a factual-knowledge test and often refuses to answer.
Artificial Analysis says Muse Ember is the efficient tier in Meta’s Muse model line. The proprietary model accepts text and images and returns text; Meta offers it through its API and its Muse Code coding agent. The evaluation lists a one-million-token context window and output speed of 440 tokens per second for the max-effort version. These are benchmark and provider-access details, not a guarantee of performance on other workloads.
The cost comparison is for an average task in Artificial Analysis’ nine-evaluation index, not a general price for every request. The source reports matching listed input and output token prices for Muse Ember and Claude Haiku 5.5, but says Ember’s cached input is cheaper: $0.004 per million cached input tokens, 60% below Haiku’s. It also says Ember used roughly a third as many output tokens as Haiku to complete the index. Together, those factors help explain the reported difference in per-task cost.
Results varied by test. Muse Ember scored 68.1% on AutomationBench-AA, level with Meta’s more expensive Muse Wonder in that evaluation, and scored higher than Haiku on Terminal-Bench 4.0. It trailed Haiku on FrontierCode 2.0 and on the AA-Briefcase knowledge-work test. In Artificial Analysis’ Coding Agent Index, the source reports a score of 46 for Ember in Muse Code, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex. Those scores reflect the tested model-and-tool combinations, not just the models in isolation.
Muse Ember scores 36 The catch behind its low task cost
Meta’s model matched Claude Haiku 5.5 on Artificial Analysis’ Intelligence Index at a lower average benchmark-task cost. The trade-off: weak factual accuracy and frequent refusals in one test.
Cost versus score
The index compares nine evaluations; its average task cost is benchmark-specific, not a universal API price.
| Model and setting | Index score | Cost per task |
|---|---|---|
| Muse Ember, max | 36 | $0.14 |
| Claude Haiku 5.5 | 36 | $0.41 |
| Muse Ember, xhigh | 34 | $0.12 |
Why the task cost is lower
The reported gap reflects usage and caching in the benchmark, despite matching listed input and output token prices.
Ember used roughly a third as many output tokens as Haiku to complete the index.
Cached input was listed at $0.004 per million tokens, 60% below Haiku’s.
Results depend on the test
Performance varied across coding, automation and knowledge-work evaluations.
Strong on one test
Ember scored 68.1% on AutomationBench-AA, level with Muse Wonder in that evaluation.
Mixed comparisons
It scored above Haiku on Terminal-Bench 4.0, but trailed on FrontierCode 2.0.
Factual caution
The cited test found low factual accuracy and frequent abstention; results may differ on your data.
What businesses should test
A low benchmark cost does not establish savings or reliability in a real workflow.
- Measure answer quality and refusal rates on customer-specific tasks.
- Track total API cost with realistic prompts, caching and output sizes.
- Decide when an abstention should trigger retrieval, a fallback model or human review.
Where Lower Inference Costs Could Help
The reported cost gap could matter to teams running high-volume, repeatable tasks, such as extracting fields, tagging records or handling simple steps in an agent workflow. At the benchmark’s stated average, $0.14 per task is about one-third of Haiku 5.5’s $0.41. Actual costs will depend on prompt length, caching, output size, reasoning settings and how a system uses the model, so the index does not establish that a business will save the same proportion.
The results also point to a potential operational trade-off: low factual accuracy and frequent abstention. Artificial Analysis reports that Ember answered 21.1% of AA-Omniscience questions correctly and declined 64% of them. A system that declines may be easier to route to a human or another model than one that provides an unsupported answer, but the benchmark does not by itself establish reliability or safety in a particular deployment. Teams would need to test both answer quality and refusal handling on their own data.
Products Worth Considering
Apple MacBook Pro 14-inch (2021) – M1 Pro Chip, 10-Core CPU, 16-Core GPU, 16GB RAM, 1TB SSD, Space Gray
Apple MacBook Pro 16.2-inch with M5 Pro, 18-core CPU, 20-core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Pro 16-inch (2021), M1 Pro, 10-core CPU, 16-core GPU, 16GB RAM, 512GB SSD, Silver
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares Models
Artificial Analysis says its Intelligence Index combines nine evaluations into a score intended to compare model performance across tasks. Its model pages list two Muse Ember reasoning settings: max scored 36 and xhigh scored 34. The reported $0.14 cost per task applies to max; xhigh scored 34 at $0.12 per task in the supplied data. These results depend on the benchmark’s tasks, model settings and cost methodology.
The same comparison places Muse Wonder at score 44 and $6.95 per index task, while Claude Fable 5.2 at its highest effort scored 65 at $26.54 per task. These figures illustrate that higher benchmark scores can come with substantially higher measured costs, but they do not establish that one model is the best choice for every use. Artificial Analysis’ comparison is dated November 9 and 10, 2026, shortly after Meta’s November 9 release.
What the Tests Do Not Establish
The supplied report does not provide independent replication, confidence intervals or enough methodological detail to determine how much the scores might vary across runs. It also does not show how the benchmark’s average task cost translates to a particular customer’s bill. API use, caching rates, prompt and response lengths, and selected reasoning effort can change costs.
The factual test results also do not establish performance across all kinds of knowledge work. AA-Omniscience measures a particular set of hard factual questions, and a refusal can be appropriate in one setting but disruptive in another. The source gives no full account of how often Ember’s refusals were correct, how the tests handle abstentions in the accuracy figure, or how the model performs on a reader’s own documents and workflows. Those points remain open pending further testing.
Testing on Real Workloads
The next useful measure will be how Muse Ember performs on customer-specific tasks, including factual accuracy, refusal rates, latency and total API cost under realistic caching and output patterns. Businesses considering it can compare those results against their existing model and define when an abstention should trigger retrieval, a fallback model or human review.
Meta’s API and Muse Code are the stated access routes in the source material. The report does not give a schedule for further benchmark updates or independent evaluations. Until those comparisons are available, the reported score and cost should be read as a snapshot of Artificial Analysis’ tests from November 2026 rather than a settled verdict on the model’s performance in production.
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 in the supplied results.
How does its reported task cost compare with Claude Haiku 5.5?
The reported average cost for one Intelligence Index task was $0.14 for Muse Ember (max) and $0.41 for Claude Haiku 5.5 (max). This is a benchmark-specific average, not a fixed cost for every API request.
Why was Muse Ember cheaper per benchmark task?
According to the source’s account of Artificial Analysis’ data, Muse Ember has lower cached-input pricing and used roughly a third as many output tokens as Haiku 5.5 to complete the index. Its listed input and output token prices were reported as matching Haiku’s.
What is the main trade-off in the evaluation?
Artificial Analysis reported 21.1% correct answers on AA-Omniscience and a 64% refusal rate for Muse Ember. Its measured hallucination rate was 17.4% in that test. These findings describe the cited benchmark and do not guarantee behavior in other settings.
Where can Muse Ember be used?
The source says Meta offers the proprietary model through Meta’s API and Muse Code. It accepts text and images and generates text, according to the supplied report.
Source: ThorstenMeyerAI.com
