🔍 Read the full analysis: Meta Muse Ember Has A 36 Score—but How Often Does It Say “I Don’t Know”? on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Meta released Muse Ember on November 9, 2026. Artificial Analysis reports that its max-effort model scores 36 on the Intelligence Index at an average cost of $0.14 per index task, matching Claude Haiku 5.5’s score at $0.41. The same evaluation finds Muse Ember answers only 21.1% of AA-Omniscience questions correctly, but declines 64% and has a reported hallucination rate of 17.4%.
Meta released Muse Ember on November 9, and Artificial Analysis reports that its max-effort version scores 36 on the Intelligence Index at an average cost of $0.14 per index task. That matches Claude Haiku 5.5’s score of 36, while the index lists Haiku’s task cost at $0.41; a separate knowledge test found Muse Ember often declines questions rather than answering them.
The figures come from Artificial Analysis, which says its Intelligence Index covers nine evaluations. Its model-page data, accessed November 10, lists Muse Ember (max) at 36 and Muse Ember (xhigh) at 34. The reported cost per task is the average API cost of completing one task in that index, not a universal price for every customer workload.
Artificial Analysis lists Muse Ember’s token prices as $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004 per million tokens. It reports output speed of 440 tokens per second and 3.2 minutes per index task for the max setting. The model accepts text and images and produces text; Meta offers it through its API and Muse Code.
The cost difference is not simply a difference in listed token rates: Artificial Analysis attributes it partly to cheaper cached input and to Muse Ember using roughly a third as many output tokens as Haiku 5.5 across the index. Those are benchmark-specific explanations. The source also reports mixed task results: Muse Ember scored 68.1% on AutomationBench-AA, while it trailed Haiku on the FrontierCode 2.0 coding test, 27.2% to 34.9%.
Muse Ember’s benchmark trade-off Lower cost, frequent refusals
Meta’s Muse Ember matches Claude Haiku 5.5 on one benchmark score at a lower reported task cost. In a separate knowledge test, it declined nearly two-thirds of questions.
Same index score, different cost
Artificial Analysis reports equal scores for Muse Ember and Claude Haiku 5.5, but different average costs under its benchmark conditions.
| Model | Index score | Cost per task |
|---|---|---|
| Muse Ember (max) | 36 | $0.14 |
| Claude Haiku 5.5 (max) | 36 | $0.41 |
| GPT-6 Luna | 28 | $0.15 |
| Muse Wonder | 44 | $6.95 |
A knowledge-test trade-off
The AA-Omniscience results show frequent declines alongside limited accuracy; its reported hallucination rate is specific to this evaluation.
Results vary by task
Muse Ember leads on some reported evaluations and trails on others, so one index score cannot predict every workload.
- AutomationBench-AA: 68.1%
- Terminal-Bench 4.0: 40.9% vs Haiku’s 32.8%
- GDP.pdf: 22.8% vs Haiku’s 20.8%
- FrontierCode 2.0: 27.2% vs Haiku’s 34.9%
- AA-Briefcase: 1,521 Elo vs Haiku’s 1,574
- AA-Omniscience accuracy: 21.1% vs Haiku’s 34.7%
What teams should test
Benchmark costs and refusal rates may not carry over to a company’s own prompts, context lengths, or usage patterns.
Test prompts and data that reflect the actual workflow.
Measure correct completions, declines, retries, and human review.
Include fallback models and review effort, not just benchmark task cost.
A Lower-Cost Option With More Refusals
The result matters to organizations running high-volume, repeatable tasks, where a lower average cost could affect the economics of extraction, tagging, form filling or agent workflows. Artificial Analysis’s comparison suggests Muse Ember can reach the same index score as Haiku 5.5 at a lower cost under its test conditions. It does not establish that the model will be cheaper or perform as well on any particular company’s workload.
The knowledge-test results point to a trade-off rather than an unqualified advantage. Artificial Analysis reports 21.1% accuracy on AA-Omniscience, alongside a 64% refusal rate and a 17.4% hallucination rate. A system that declines uncertain questions may be easier to route to a human or a second model, but refusals can also interrupt workflows. Buyers would need to test both answer quality and refusal handling on their own data.
For practical use, the figures favor tasks where a model receives the relevant facts in documents, retrieved material or databases, rather than being expected to recall obscure information. That is an interpretation of the reported results, not a guarantee about performance in deployed systems.
Popular right now
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
KitchenAid Classic Series 4.5-Quart Tilt-Head Stand Mixer, Onyx Black
Shark NV752 Rotator Powered Lift-Away TruePet Upright Vacuum, Bordeaux
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares Models
Muse Ember is described in the source as the efficient tier of Meta’s Muse model line. Artificial Analysis tested two reasoning-effort settings, max and xhigh. On its reported comparison, Muse Ember (max) and Claude Haiku 5.5 (max) each score 36; their respective index-task costs are $0.14 and $0.41. The same table lists GPT-6 Luna at 28 and $0.15, and Meta’s Muse Wonder at 44 and $6.95.
Artificial Analysis’s AA-Omniscience evaluation allows models to decline questions. In that test, the source reports Muse Ember’s accuracy at 21.1%, Haiku 5.5’s at 34.7%, and Gemini 3.8 Flash (high)’s at 50.5%. It gives Muse Ember a 17.4% hallucination rate, compared with 34.2% for Haiku 5.5 and 79.3% for GPT-6 Luna. These are results from that particular test, not general error rates across all uses.
Other reported scores vary by task. Muse Ember reached 22.8% on GDP.pdf, slightly above Haiku’s 20.8%, but scored lower on the AA-Briefcase knowledge-work test, at 1,521 Elo versus 1,574. On Terminal-Bench 4.0 it scored 40.9% against Haiku’s 32.8%. The source also reports a score of 46 for Muse Ember in Meta’s Muse Code on the Artificial Analysis Coding Agent Index; comparisons across different model-and-tool setups should be read as benchmark results, not direct predictions of coding quality for every user.
What the Test Results Cannot Show
The figures are benchmark results reported by Artificial Analysis; the source material does not provide independent replication or enough methodological detail to verify the comparisons here. It is not clear how closely the index tasks match real customer workloads, or whether the reported average cost will hold across different prompts, context lengths, account terms and usage patterns.
The source does not give a complete breakdown of refusal rates across ordinary business tasks, nor does it show how often a refusal would be appropriate or how users respond to one. The reported 17.4% hallucination rate applies to AA-Omniscience as described, and should not be treated as a rate for all questions or deployments. Pricing and model availability may also change after the November 9 release.
Testing on Real Workloads
The next useful comparison is whether Muse Ember’s reported cost and performance carry over to tasks customers actually run. Teams considering it can test representative prompts, measure the rate of correct completions and refusals, and account for the cost of retries, human review or fallback models. The source recommends measuring on a customer’s own tasks rather than relying on the index alone.
Meta’s API and Muse Code are the cited access routes, but the source does not specify a broader rollout schedule or future model updates. Further independently published tests and clearer deployment-specific pricing would help establish how the benchmark trade-off translates into operational use.
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. Its xhigh setting scored 34 in the cited model-page data.
How does Muse Ember’s reported task cost compare with Haiku 5.5?
For one task in the Artificial Analysis Intelligence Index, the source reports an average cost of $0.14 for Muse Ember (max) and $0.41 for Claude Haiku 5.5 (max). These are benchmark costs, not guaranteed prices for other workloads.
How often does Muse Ember say it does not know?
Artificial Analysis reports that Muse Ember declined 64% of questions in its AA-Omniscience test. That rate applies to the cited evaluation and may differ on other prompts or tasks.
Does Muse Ember perform better than Haiku 5.5 at every task?
No. The reported results are mixed: Muse Ember scored higher on some evaluations, including AutomationBench-AA and Terminal-Bench 4.0, but lower on FrontierCode 2.0 and the AA-Briefcase test. Benchmark outcomes do not establish which model will perform better on a specific user’s workload.
Source: ThorstenMeyerAI.com
