TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meta released Muse Ember on Nov. 9, and Artificial Analysis rates its max-effort version at 36 on its Intelligence Index, matching Claude Haiku 5.5. The evaluator estimates $0.14 per index task for Muse Ember versus $0.41 for Haiku, while reporting lower factual accuracy and a higher rate of refusals on its knowledge test.
Meta released Muse Ember on Nov. 9, and Artificial Analysis says the model’s max-effort version scored 36 on its nine-evaluation Intelligence Index, matching Claude Haiku 5.5. The evaluator estimates an average cost of $0.14 per index task for Ember, compared with $0.41 for Haiku, while its separate factual-knowledge results show lower accuracy and more refusals.
Artificial Analysis lists Muse Ember as a proprietary Meta model available through Meta’s API and the Muse Code coding agent. It accepts text and images and generates text. The evaluator tested two reasoning effort settings: max scored 36 at an estimated $0.14 per index task; xhigh scored 34 at $0.12. The xhigh run was faster in the index, at 2.8 minutes per task compared with 3.2 minutes for max, according to the model data accessed Nov. 10.
The listed token prices are $0.10 per million input tokens and $0.50 per million output tokens, the same rates Artificial Analysis gives for Claude Haiku 5.5 and GPT-6 Luna. Cached input is priced at $0.004 per million tokens, which the evaluator says is 60% below Haiku’s cached-input price. Artificial Analysis attributes Ember’s lower task cost partly to that cache rate and partly to its use of roughly one-third as many output tokens as Haiku in the index. Those cost estimates describe the evaluator’s benchmark workload, not a typical customer’s bill.
Results vary by task. Ember scored 68.1% on AutomationBench-AA, matching the more expensive Muse Wonder result cited by the source and exceeding Haiku’s 35.4%. It scored 22.8% on GDP.pdf, compared with Haiku’s 20.8%, but recorded a lower score on the AA-Briefcase knowledge-work test: 1,521 Elo to Haiku’s 1,574. On coding tests, Ember led Haiku on Terminal-Bench 4.0, 40.9% to 32.8%, and trailed it on FrontierCode 2.0, 27.2% to 34.9%.
Meta’s Muse Ember: A Score of 36, Cheap Tasks, Plenty of Doubt
Released Nov. 9, Muse Ember pairs a competitive benchmark score with low estimated task costs. Its results vary by task, and factual answers often give way to refusals.
A lower-cost alternative
Artificial Analysis estimates Ember’s index-task cost at about a third of Claude Haiku 5.5’s.
- $0.14 per index task at max effort
- $0.12 at xhigh effort
- Lower cost partly reflects cached input and fewer output tokens
- $0.41 per index task
- Same listed input and output token prices
- Higher estimated cost on the evaluator’s workload
Strengths depend on the task
Benchmark results point to useful wins, but no single score shows which model fits a company’s workflow.
Strong result
Ember scored 68.1% on AutomationBench-AA, above Haiku’s 35.4%.
Mixed comparison
Ember led Haiku on Terminal-Bench 4.0, but trailed on FrontierCode 2.0.
A weaker showing
Ember scored 1,521 Elo on AA-Briefcase, below Haiku’s 1,574.
Factual answers come with caveats
On AA-Omniscience, Ember’s lower reported hallucination rate coincided with frequent refusals.
What businesses should test
The published index is a comparison point, not a forecast of production performance or cost.
Check performance on the tasks and tools your team actually uses.
Decide whether Ember’s tendency to decline is acceptable for each workflow.
Account for prompt length, cache reuse, output needs, settings and correction work.
Lower Costs, Uneven Results
The reported price gap could matter for companies that run high volumes of repetitive model calls. At the evaluator’s estimated prices, one index task with Ember costs about a third of one with Haiku. If a company’s work resembles the benchmark tasks, the difference could affect the cost of routing, extraction, tagging, form filling or repeated agent steps. The published figures do not establish what those workloads would cost in production; task length, cached context, model settings and error handling can change the calculation.
Performance also makes model choice a task-specific decision. Ember’s AutomationBench-AA result is strong in the comparison provided, while its lower AA-Briefcase score and mixed coding results show that a single index score cannot establish which model will perform best for a given company. Teams would need to compare models on their own examples and account for the cost of checking or correcting outputs.
The knowledge-test results point to another trade-off. Artificial Analysis reports that Ember answers correctly less often than the compared models but declines many questions, limiting its reported hallucination rate. Refusing to guess can be useful where an unsupported factual answer creates risk, but refusals can also interrupt work that expects a completed answer. The benchmark does not show how the model behaves across every subject, prompt style or deployment setting.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
Apple AirPods Pro (1st Generation) with MagSafe Charging Case
As an affiliate, we earn on qualifying purchases.
How Ember Compares
Muse Ember is presented as the efficient tier of Meta’s Muse model line. In the cited Artificial Analysis data, Meta’s larger Muse Wonder scores 44 on the Intelligence Index and costs an estimated $6.95 per index task, compared with Ember’s 36 and $0.14. The source also gives Claude Fable 5.2 a score of 65 at its highest effort setting, with an estimated task cost of $26.54. These are evaluator estimates from its index, not general price quotes for all uses.
The Intelligence Index combines nine evaluations, so its score compresses varied capabilities into one number. The source reports that Ember generates 440 output tokens per second through Meta’s API and has a one-million-token context window. It also reports a score of 46 for Ember in the Artificial Analysis Coding Agent Index when used through Muse Code; the listed comparisons are 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex. These agent-index results depend on the named tools and evaluation setup.
Artificial Analysis’s AA-Omniscience test asks difficult factual questions and allows a model to decline. In that test, the source reports that Ember answered 21.1% correctly and declined 64% of questions, with a 17.4% hallucination rate. Haiku answered 34.7% correctly and had a 34.2% hallucination rate; GPT-6 Luna’s reported rate was 79.3%. These figures describe the test’s results and should not be read as universal error rates across everyday use.
Limits of the Benchmark Evidence
The published comparison does not establish how Muse Ember will perform on an individual organization’s work. Artificial Analysis’s index tasks and test suites provide a basis for comparison, but the source itself advises measuring performance on a company’s own tasks. The article’s source material does not provide independent production results, customer deployment data or a detailed breakdown of how benchmark outcomes vary across industries.
It is also unclear from the available material how the reported costs would change with different prompt lengths, cache reuse, output requirements or reasoning settings outside the tested runs. The source describes Muse Ember as proprietary and available through Meta’s API and Muse Code, but does not give further detail here on access terms or deployment limits. Those details may affect whether the benchmark’s low per-task estimate translates into lower operating costs.
Testing Ember in Real Workflows
The next practical measure is how Muse Ember performs in customer workloads and direct comparisons against alternatives. Businesses considering it can test representative prompts, track correct answers and refusals, and calculate costs under their expected usage patterns. For factual tasks, those checks can show whether Ember’s tendency to decline is acceptable; for automation or coding, they can reveal whether benchmark strengths carry over to the tools and workflows a team uses.
Artificial Analysis’s data was accessed Nov. 10, one day after Meta’s release. The source material provides no later release milestone or scheduled update. Further independent testing, along with details from Meta about access and deployment, would add evidence about how broadly the benchmark findings apply.
Key Questions
What did Muse Ember score on the Intelligence Index?
Artificial Analysis scored Muse Ember at 36 at max effort and 34 at xhigh. The index combines nine evaluations.
How much does a Muse Ember index task cost?
Artificial Analysis estimates $0.14 per task for the max-effort run and $0.12 for xhigh. These are average costs for its index workload, not a guaranteed cost for other tasks.
Does Muse Ember answer factual questions reliably?
In Artificial Analysis’s AA-Omniscience test, Ember answered 21.1% correctly and declined 64% of questions. Its reported hallucination rate in that test was 17.4%; results may differ on other prompts and subject areas.
Where did Muse Ember perform well?
It scored 68.1% on AutomationBench-AA and outscored Haiku on Terminal-Bench 4.0 in the comparison reported. It scored below Haiku on FrontierCode 2.0 and AA-Briefcase.
Source: Thorsten Meyer AI
