🔍 Read the full analysis: Meta’s Muse Ember Balances A 36 Score With Frequent “I Don’t Know” Replies 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 November 9, 2026. Artificial Analysis gives its max-effort version an Intelligence Index score of 36 and an average cost of $0.14 per index task, compared with $0.41 for Claude Haiku 5.5 at the same score. The benchmark also found lower factual accuracy but more refusals and a lower hallucination rate; results may not predict performance on every workload.
Meta released Muse Ember on November 9, and independent benchmark provider Artificial Analysis scored its max-effort version at 36 on the Intelligence Index, matching Claude Haiku 5.5 while costing $0.14 per index task compared with $0.41. The results point to a trade-off for businesses considering the model: it is cheaper on this benchmark, but answered fewer factual questions correctly and often declined to answer.
Artificial Analysis says the Intelligence Index combines nine evaluations. Its reported average API cost per task was $0.14 for Muse Ember at max reasoning effort and $0.41 for Claude Haiku 5.5 at max, despite both models receiving a score of 36. The figures refer to the benchmark’s tasks, not a guaranteed cost for a particular company’s workload.
The model’s listed prices are $0.10 per million input tokens and $0.50 per million output tokens, according to the source. Artificial Analysis attributes its lower index-task cost in part to cached input costing $0.004 and to Muse Ember using roughly a third as many output tokens as Haiku 5.5 across the index. The source reports an output speed of 440 tokens per second, a 3.2-minute average per index task, and a one-million-token context window for the max version.
Results varied by evaluation. Muse Ember scored 68.1% on AutomationBench-AA, matching Meta’s more expensive Muse Wonder and exceeding Haiku 5.5’s 35.4% in the reported comparison. On coding tests, it scored higher than Haiku on Terminal-Bench 4.0 but lower on FrontierCode 2.0. Artificial Analysis also reported a Coding Agent Index score of 46 for Muse Ember in Meta’s Muse Code tool, compared with 37 for Haiku 5.5 in Claude Code and 31 for GPT-6 Luna in Codex.
Meta’s Muse Ember costs less, declines more
Artificial Analysis gave Muse Ember and Claude Haiku 5.5 the same Intelligence Index score. Its benchmark also found different costs and factual-answer trade-offs.
Benchmark cost and score
The reported task costs apply to Artificial Analysis’s index, not every company workload.
| Model | Index score | Cost per task |
|---|---|---|
| Muse Ember, max effort | 36 | $0.14 |
| Claude Haiku 5.5, max | 36 | $0.41 |
Factual answers involve a trade-off
On AA-Omniscience, Ember declined many questions and hallucinated less, but answered fewer correctly.
- 21.1% factual accuracy
- 64% of questions declined
- 17.4% reported hallucination rate
- 34.7% factual accuracy
- 34.2% reported hallucination rate
Results vary by task
The benchmark does not establish how either model will perform on a company’s own data.
Strong on one test
Ember scored 68.1% on AutomationBench-AA, matching Muse Wonder and exceeding Haiku 5.5’s 35.4% in the comparison.
Mixed comparison
Ember scored higher than Haiku on Terminal-Bench 4.0 but lower on FrontierCode 2.0.
What businesses should assess
Test representative work before choosing where Ember fits.
- Measure accuracy, refusal frequency, latency, and total cost.
- Check performance on your own prompts and data.
- Plan whether declined questions go to a person or another model.
Lower Costs, More Cautious Answers
The headline comparison may matter to teams running high-volume, repeated tasks, where per-task costs can add up. The benchmark’s $0.14 versus $0.41 result suggests Muse Ember could be worth testing for workflows such as extraction, tagging, form filling, or agent steps. It does not establish that the model will be cheaper or perform as well on every company’s data: actual costs depend on prompts, context, output length, and usage patterns.
Its factual-question performance presents a separate operational trade-off. On Artificial Analysis’s AA-Omniscience test, Muse Ember answered 21.1% correctly, declined 64% of questions, and had a reported hallucination rate of 17.4%. Haiku 5.5 scored 34.7% accuracy and had a 34.2% hallucination rate in the same comparison. A high refusal rate can reduce unsupported answers, but it can also mean more questions need human review or another model. The figures describe this test, not every use of the models.
For organizations, the results support evaluating the model where source material is provided directly, such as document processing or tasks grounded in a company database. They also suggest that systems should define what happens when the model declines. A lower benchmark cost is useful only if the workflow can handle refusals and the model performs adequately on the organization’s own examples.
Products Worth Considering
Apple Mac mini (2026), M6 12-core CPU and GPU, 24GB Memory, 1TB SSD
Apple Mac mini Desktop Computer with M2 Pro, 16GB Memory, 512GB SSD, Gigabit Ethernet
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
As an affiliate, we earn on qualifying purchases.
What Artificial Analysis Tested
Muse Ember is described in the source as the efficient tier of Meta’s Muse model line, a proprietary model available through Meta’s API and in Muse Code. It accepts text and images as input and produces text. Artificial Analysis tested two reasoning-effort settings: max and xhigh. The max version scored 36; xhigh scored 34 and had a reported index-task cost of $0.12.
Artificial Analysis’s comparison places Muse Ember’s $0.14 task cost below Haiku 5.5’s $0.41, but the source also lists substantial differences among models with higher scores. Meta’s Muse Wonder scored 44 at a reported $6.95 per index task, while Claude Fable 5.2 scored 65 at $26.54 at its highest effort setting. These are benchmark costs and scores, not a universal ranking of quality or a promise about production pricing.
The source’s task-level explanation is that cached context accounts for much of the cost in agentic evaluations, and Muse Ember’s cached input price is lower than Haiku’s. It also says Ember used fewer output tokens during the index. That helps explain how two models with similar listed per-token prices can have different average costs in this particular test.
Limits of the Benchmark Results
The available source does not provide enough detail to determine how Muse Ember will perform across specific customer workloads, or whether the benchmark’s average cost will carry over to production. The Intelligence Index is one set of evaluations, and the source itself advises testing models on a team’s own tasks before choosing a routing or deployment strategy.
It is also unclear from the supplied material how refusal behavior varies by prompt, subject, or model version, and what fallback behavior Meta recommends when Muse Ember says it does not know. The benchmark reports scores and rates, but does not establish why the model declined particular questions or whether its answers are reliable in every domain. Pricing, access terms, and future model changes could also affect a business comparison.
Testing Ember in Real Workflows
The next practical step for prospective users is to compare Muse Ember with alternatives on representative tasks, measuring accuracy, refusal frequency, latency, and total cost rather than relying on an index score alone. For tasks where the model can be given the relevant facts, teams can check whether its performance and lower reported benchmark cost justify a trial.
Organizations considering deployment will also need a plan for unanswered questions, such as routing refusals to a human or another model. The supplied source does not identify a future benchmark release or a specific Meta roadmap milestone. Further comparisons will depend on new test data, any changes to Meta’s model or pricing, and results from users’ own workloads.
Key Questions
What did Meta release?
Meta released Muse Ember on November 9, 2026. The source describes it as the efficient tier of its Muse model line, available through Meta’s API and in Muse Code.
How did Muse Ember score against Claude Haiku 5.5?
Artificial Analysis gave both max-effort versions an Intelligence Index score of 36. It reported an average cost of $0.14 per index task for Muse Ember and $0.41 for Haiku 5.5.
Why does Muse Ember often say it does not know?
In Artificial Analysis’s AA-Omniscience factual test, Muse Ember declined 64% of questions. The supplied material reports that result but does not explain the model’s decision process for individual questions.
Does a lower benchmark cost mean Muse Ember will be cheaper for every business?
No. The $0.14 figure is an average cost per task in Artificial Analysis’s index. A company’s cost and results can differ depending on its prompts, data, context length, output needs, and workload.
What are Muse Ember’s reported weaknesses?
Artificial Analysis reported 21.1% accuracy on AA-Omniscience and mixed coding results: Muse Ember beat Haiku 5.5 on Terminal-Bench 4.0 but trailed it on FrontierCode 2.0. Those findings apply to the named evaluations.
Source: ThorstenMeyerAI.com
