🔍 Read the full analysis: Meta’s Muse Ember Scores 36—But Often Won’t Give An Answer on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Meta released Muse Ember on Nov. 9, 2026. Artificial Analysis gave its max-effort version an Intelligence Index score of 36, matching Claude Haiku 5.5, while estimating a lower cost per index task: $0.14 versus $0.41. The tests also found low factual accuracy alongside frequent refusals and a lower hallucination rate than several compared models.
Meta released Muse Ember on Nov. 9, and independent benchmark data from Artificial Analysis put its max-effort version at an Intelligence Index score of 36—the same as Claude Haiku 5.5. Artificial Analysis estimated that a task in its index cost $0.14 with Muse Ember, compared with $0.41 for Haiku, while also finding that Ember answers relatively few questions in a factual-knowledge test and often declines to respond.
Muse Ember is the efficient tier of Meta’s Muse model line, a proprietary model available through Meta’s API and the company’s Muse Code coding agent. It accepts text and images and produces text. Artificial Analysis tested it at two reasoning-effort settings. At max effort, it scored 36; at xhigh, it scored 34. The reported task cost was $0.14 at max and $0.12 at xhigh.
The cost figures refer to the average API cost of one task in Artificial Analysis’s nine-evaluation index, not a universal price for every customer request. The source says Muse Ember’s listed token prices match Haiku 5.5 and GPT-6 Luna: $0.10 per million input tokens and $0.50 per million output tokens. Cached input is listed at $0.004 per million tokens, which the source describes as 60% less than Haiku’s. Artificial Analysis also reported 440 output tokens per second and 3.2 minutes per index task for Ember at max effort.
The benchmark results vary by task. Muse Ember scored 68.1% on AutomationBench-AA, level with Meta’s more expensive Muse Wonder and above Haiku 5.5’s 35.4% in the comparison. On coding tests, Ember scored higher than Haiku on Terminal-Bench 4.0, 40.9% to 32.8%, but lower on FrontierCode 2.0, 27.2% to 34.9%. In the Artificial Analysis Coding Agent Index, it scored 46 when used through Muse Code; the source lists Haiku in Claude Code at 37 and GPT-6 Luna in Codex at 31.
Muse Ember scores 36 but often won’t answer
Meta’s model matches Claude Haiku 5.5 on one benchmark at a lower estimated task cost, but factual accuracy is low and refusals are frequent.
A benchmark cost advantage
The index score matches Haiku 5.5, while the estimated cost per task is lower; this is not a guaranteed saving on every workload.
| Model | Index score | Cost per task |
|---|---|---|
| Muse Ember (max) | 36 | $0.14 |
| Claude Haiku 5.5 (max) | 36 | $0.41 |
| GPT-6 Luna (max) | 28 | $0.15 |
| Muse Wonder (max) | 44 | $6.95 |
Factual answers come with a trade-off
On AA-Omniscience, Ember often abstained and had lower factual accuracy than the reported Haiku comparison.
Performance varies by task
Benchmark results point to strengths in some workflows, not an overall win across coding and business tasks.
- AutomationBench-AA: 68.1%, level with Muse Wonder and above Haiku’s 35.4%.
- Terminal-Bench 4.0: 40.9% versus Haiku’s 32.8%.
- FrontierCode 2.0: 27.2% versus Haiku’s 34.9%.
- Results depend on the task and, for coding, the agent used.
Where it may fit
A lower benchmark cost may suit repeatable work grounded in supplied documents or retrieved information, provided teams manage unanswered requests.
Compare Ember with alternatives on your prompts, data, and answer-quality needs.
Include request costs, retries, fallbacks, review, and latency targets.
Decide when retrieval, another model, or human review should handle unanswered requests.
Lower Cost, With a Knowledge Trade-Off
The results put Muse Ember in a potentially useful position for high-volume, repeatable workflows where the system works from supplied documents or retrieved information. If its benchmark cost advantage carries over to a company’s own use, tasks such as extraction, tagging, form completion and basic agent steps could cost less than with a similarly scoring alternative. The index estimate alone does not establish savings for a particular workload.
Its tendency to decline factual questions changes how it may fit into products. Artificial Analysis’s AA-Omniscience test found 21.1% accuracy and a 64% refusal rate for Ember, with a 17.4% hallucination rate. Those results suggest a model that may be more likely to abstain than to supply an unsupported factual answer. That can be useful where wrong answers carry risk, but can also mean more unanswered requests and a need for fallback systems or human review.
For buyers, the takeaway is not simply that a lower benchmark cost makes Ember the better choice. Results differ by task, and the source reports both areas of strength and weaker performance. Teams would need to test it on their own prompts, data, refusal requirements and latency targets before making deployment decisions.
Popular right now
Vitamix 5200 Blender with 64 oz. Container, Black
Samsung 43-Inch M70H Series 4K Mini LED Smart TV (2026)
LG 77-Inch OLED evo C6 4K Smart TV
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares Models
Artificial Analysis’s Intelligence Index combines results from nine evaluations; its score is a benchmark measure, not a guarantee of performance in a specific application. The source’s comparison lists Muse Ember (max) and Claude Haiku 5.5 (max) at 36, while GPT-6 Luna (max) scores 28 and Meta’s Muse Wonder (max) scores 44. Their reported costs per index task are $0.14, $0.41, $0.15 and $6.95, respectively.
The source attributes Ember’s lower task cost relative to Haiku, despite matching listed token prices, to two factors: its cached input is cheaper and it used roughly a third as many output tokens as Haiku to complete the index. These are explanations reported in the source based on Artificial Analysis data. They do not mean every task will produce the same cost difference.
The knowledge comparison adds another dimension to the score. On AA-Omniscience, Artificial Analysis reported 21.1% accuracy for Muse Ember, 34.7% for Haiku 5.5 and 50.5% for Gemini 3.8 Flash at high effort. Ember’s refusal rate was 64%, while its hallucination rate was 17.4%; the source lists Haiku’s hallucination rate at 34.2% and GPT-6 Luna’s at 79.3%. Those figures describe performance on that specific test, not all factual questions.
Limits of the Reported Results
The figures available here come from Artificial Analysis benchmarks and model-page data accessed on Nov. 9 and 10, 2026. The source does not provide enough detail to establish how the scores translate to individual business tasks, different prompts or production settings. Nor does the index cost estimate show what every user would pay after accounting for request size, caching, service terms and usage patterns.
The material also does not specify how often Muse Ember’s refusals occur outside AA-Omniscience or how readily a product could resolve them through retrieval, a second model or human review. The reported results show mixed performance across coding and document tasks, but do not explain all differences between evaluations. Claims about cost and reliability should therefore be read as benchmark findings, not broad guarantees.
Testing Ember in Real Workflows
The next useful measure will be performance on customer workloads: whether the lower index-task cost holds for routine requests, how often refusals interrupt a process, and whether supplied documents or retrieved data improve factual results. Teams considering the model can compare it with alternatives on representative tasks, including both answer quality and the cost of retries, fallbacks and review.
Meta’s release makes Muse Ember available through its API and Muse Code, according to the source. No further release date, pricing change or additional benchmark schedule is specified in the material. For now, the confirmed development is the model’s launch and its reported benchmark profile; wider adoption and real-world performance remain to be established.
Key Questions
What is Meta Muse Ember?
Muse Ember is the efficient tier of Meta’s proprietary Muse model line. The source says it accepts text and images and is available through Meta’s API and Muse Code.
How did Muse Ember score against Claude Haiku 5.5?
At max reasoning effort, both scored 36 on Artificial Analysis’s Intelligence Index. The source estimates an index task cost of $0.14 for Ember and $0.41 for Haiku; that is a benchmark-specific estimate, not a guaranteed cost for other workloads.
Why does Muse Ember often refuse to answer?
Artificial Analysis reported that Ember declined 64% of questions on AA-Omniscience, a factual-knowledge test. The source does not establish that the same refusal rate applies to other tasks or real-world use.
Is Muse Ember better at coding than Haiku 5.5?
The results are mixed. Ember scored higher on Terminal-Bench 4.0 but lower on FrontierCode 2.0 in the reported comparison. Its score also depended on the coding agent used, so the benchmark results do not establish an overall winner for coding.
Source: ThorstenMeyerAI.com
