🔍 Read the full analysis: Why Meta’s Muse Ember’s 36 Score Doesn’t Tell The Whole Story 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 scores its max-effort version 36, matching Claude Haiku 5.5, at an estimated $0.14 per index task versus $0.41 for Haiku; its tests also show weaker factual accuracy but more frequent refusals and a lower hallucination rate.
Meta released Muse Ember on November 9, and an Artificial Analysis evaluation gives its max-effort version an Intelligence Index score of 36—the same as Claude Haiku 5.5—at an estimated $0.14 per index task, compared with $0.41 for Haiku. The results point to a lower-cost option for some repeated AI work, but the benchmark also finds weaker factual accuracy and frequent refusals, so the score alone does not establish which model will perform better for a particular business.
Artificial Analysis says the Intelligence Index combines nine evaluations. Its model pages list Muse Ember (max) at 36 and $0.14 per index task. Muse Ember (xhigh), another tested reasoning setting, scores 34 at $0.12 per task. These are benchmark cost estimates, not a quoted price for every real-world request. Artificial Analysis’s figures show max-effort Muse Ember producing 440 output tokens per second and taking 3.2 minutes to complete the index tasks; the model has a one-million-token context window.
The model’s listed prices are $0.10 per million input tokens and $0.50 per million output tokens, matching the listed rates for Claude Haiku 5.5 and GPT-6 Luna. Artificial Analysis attributes Ember’s lower index-task cost relative to Haiku to its cached input rate, which it says is 60% lower, and to Ember using roughly a third as many output tokens to complete the index. Those explanations are based on the evaluation’s workload and should not be treated as a guarantee of savings for other tasks.
Performance varies by test. Artificial Analysis reports 68.1% on AutomationBench-AA, matching its result for Meta’s more expensive Muse Wonder, while Ember scores 27.2% on FrontierCode 2.0, below Haiku’s 34.9%. It does better than Haiku on Terminal-Bench 4.0, with 40.9% against 32.8%. In Meta’s Muse Code tool, Ember scores 46 on Artificial Analysis’s Coding Agent Index, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex.
Why Muse Ember’s 36 Score Doesn’t Tell the Whole Story
Meta’s Muse Ember matches Claude Haiku 5.5 on Artificial Analysis’s Intelligence Index, with a lower estimated cost per benchmark task. But performance, factual accuracy and refusals vary by test.
Where Ember Performs
Results shift across benchmarks, so the index score is not a verdict for every workload.
The Factuality Trade-Off
On AA-Omniscience, fewer hallucinations come alongside lower accuracy and frequent refusals.
| Model | Correct answers | Hallucination rate |
|---|---|---|
| Muse Ember | 21.1% | 17.4% |
| Claude Haiku 5.5 | 34.7% | 34.2% |
| GPT-6 Luna | Not provided | 79.3% |
Where to Test First
Lower benchmark cost may suit repeatable work, but savings depend on the full workflow.
Grounded, repeatable tasks
Try document extraction, tagging or form filling when the model can use supplied source information.
Open-ended factual work
Lower accuracy and frequent refusals may interrupt tasks or require fallback models and human review.
Before Switching
Compare models on representative tasks from your own data and operating conditions.
- Measure accuracy and task completion.
- Track refusals, retries and review needs.
- Include latency and full workflow cost.
- Test the reasoning setting you plan to use.
Lower Costs, Uneven Performance
The result matters to companies that run high-volume, repeatable workflows, where modest differences in cost per task can add up. Document extraction, tagging, form filling, and routine agent steps may be suitable places to test a cheaper model, especially when the model is given the relevant facts rather than asked to recall them. The benchmark’s price comparison gives buyers a reason to evaluate Ember, not proof that it will lower costs on their own systems.
Its refusal behavior is another practical distinction. Artificial Analysis’s AA-Omniscience test records Ember answering 21.1% correctly, declining 64% of questions, and hallucinating at a rate of 17.4%. Haiku answers 34.7% correctly and has a reported hallucination rate of 34.2% in that test. A system that often declines may be preferable in settings where unsupported answers are risky, but a refusal can still interrupt a workflow and may require a fallback model or human review.
The trade-off is not simply “cheap versus capable.” Results differ across factual knowledge, coding, and workflow tests. Teams need to compare models on their own data, including the cost of refusals, retries, and review. A score of 36 is a summary across tests, not a measure of reliability for every application.
Products Worth Considering
Bose Lifestyle Ultra Wireless Home Speaker, White Smoke
Bose Smart Soundbar 700 with Alexa, Black
Samsung Galaxy Tab S7 11-inch Android Tablet 128GB Wi‑Fi with S Pen Mystic Black
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares
Muse Ember is described in the source material as the efficient tier of Meta’s Muse model line, available through Meta’s API and the Muse Code coding agent. It accepts text and images as input and produces text. Artificial Analysis tested it at max and xhigh reasoning effort, so reported scores and costs depend in part on which setting is used.
The same benchmark data put Claude Haiku 5.5 (max) at 36 for $0.41 per task, and GPT-6 Luna (max) at 28 for $0.15. Meta’s Muse Wonder scores 44 at $6.95 per task. Artificial Analysis also reports Claude Fable 5.2 at 65 and $26.54 per task at its highest effort level. These are comparisons within Artificial Analysis’s index, not broad measures of model quality or independent assessments of every vendor’s pricing.
On AA-Omniscience, which lets models decline difficult factual questions, Ember’s reported hallucination rate is lower than those of the named comparators: 34.2% for Haiku and 79.3% for GPT-6 Luna. Its lower accuracy and high refusal rate are part of the same result; the figures should be read together rather than treating the hallucination rate on its own as a complete account of factual performance.
What the Tests Cannot Establish
The supplied results do not show how Muse Ember performs across customers’ specific data, prompts, and operating conditions. Artificial Analysis’s benchmark cost is an average for its index tasks; actual costs will depend on input and output length, caching, reasoning settings, and how often a system needs retries or human review. The source material does not provide a matched evaluation of all the models on one organization’s production workload.
It is also unclear from the available material how representative each test is of commercial use, how much performance changes under different prompts, and what safeguards or service terms accompany Meta’s API and Muse Code. The benchmark figures are attributed to Artificial Analysis; they should not be read as a guarantee from Meta or as independently verified outcomes for every deployment. No broader conclusion about overall model safety follows from the reported hallucination rate alone.
Test Before Switching Workloads
For teams considering Ember, the next step is a side-by-side trial on representative tasks. That should measure accuracy, refusal frequency, task completion, latency, and full cost—including fallback calls and review—not just token prices or index scores. Work that supplies the model with source documents or retrieved information may be a more appropriate early test than open-ended factual questions, given the reported knowledge results.
Artificial Analysis’s published results provide an initial comparison, but the supplied source does not identify a further benchmark release date or a specific next product milestone from Meta. Any adoption decision will depend on how Ember performs in real deployments and whether its lower benchmark cost offsets refusals and other workflow needs.
Key Questions
What does Muse Ember’s score of 36 mean?
It is Muse Ember (max)’s score on the Artificial Analysis Intelligence Index, which combines nine evaluations. It matches Claude Haiku 5.5’s score in the cited comparison and does not by itself predict performance on every task.
Is Muse Ember cheaper than Claude Haiku 5.5?
Artificial Analysis estimates $0.14 per index task for Muse Ember (max), compared with $0.41 for Haiku (max). The listed per-token input and output prices are the same in the source material; the task-cost difference is attributed to cheaper cached input and fewer output tokens on the benchmark.
Does Muse Ember hallucinate less?
In Artificial Analysis’s AA-Omniscience test, Ember’s reported hallucination rate is 17.4%, compared with 34.2% for Haiku and 79.3% for GPT-6 Luna. Ember also declines 64% of questions and answers 21.1% correctly, so the lower rate needs to be considered alongside its high refusal rate and lower accuracy.
Is Muse Ember a strong coding model?
Its coding results are mixed. It scores 40.9% on Terminal-Bench 4.0, ahead of Haiku’s 32.8%, but scores 27.2% on FrontierCode 2.0, below Haiku’s 34.9%. In Meta’s Muse Code tool, it scores 46 on Artificial Analysis’s Coding Agent Index.
Source: ThorstenMeyerAI.com
