🔍 Read the full analysis: The Numbers Behind Meta’s Muse Ember: Score 36 And 14 Cents Per Task 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 gives its max-effort version an Intelligence Index score of 36 and estimates an index task costs $0.14, compared with $0.41 for Claude Haiku 5.5 at the same score. The benchmark also finds weaker factual accuracy but a higher rate of refusals and a lower hallucination rate for Muse Ember; results may differ on real workloads.
Meta released Muse Ember on November 9, and an Artificial Analysis evaluation gives its max-effort version an Intelligence Index score of 36, equal to Claude Haiku 5.5’s score but with a reported average cost of $0.14 per index task versus $0.41. The comparison makes Muse Ember a lower-cost option in this benchmark, while the same evaluation shows weaker factual accuracy and more frequent refusals.
Artificial Analysis says the Intelligence Index combines nine evaluations. Its estimated cost per task is the average API cost of completing one task in that index, not a guaranteed price for a customer’s particular application. The results are specific to the tested settings: Muse Ember (max) scored 36, while its xhigh setting scored 34 at a reported $0.12 per index task. At max effort, Claude Haiku 5.5 also scored 36 and cost $0.41 per task in the comparison.
Meta’s listed rates for Muse Ember are $0.10 per million input tokens and $0.50 per million output tokens. The source says those rates match Haiku 5.5 and GPT-6 Luna, while Muse Ember’s cached input rate is $0.004 per million tokens, 60% below Haiku’s. Artificial Analysis attributes the gap in task costs to cheaper cached input and Muse Ember using roughly a third as many output tokens as Haiku 5.5 to complete the index.
The model accepts text and images and generates text. Meta offers it through its API and its Muse Code coding agent. Artificial Analysis reports output speed of 440 tokens per second and an index completion time of 3.2 minutes for the max setting, with a one-million-token context window. Those figures describe the benchmark setup and should not be treated as a promise of performance on every request.
Meta’s Muse Ember Score 36, lower benchmark cost
Artificial Analysis reports that Muse Ember matches Claude Haiku 5.5’s Intelligence Index score at a lower estimated cost per task. The benchmark also finds trade-offs in factual accuracy and refusals.
What the comparison shows
The task-cost figures are benchmark estimates, not guaranteed costs for a company’s own workload.
- Lower estimated index-task cost: $0.14
- Higher refusal rate in the factual-question test
- Lower reported hallucination rate in that test
- Estimated index-task cost: $0.41
- Same Intelligence Index score: 36
- Higher factual accuracy in the AA-Omniscience test
Performance depends on the task
Results vary across the reported evaluations; one overall score does not establish which model will fit a specific workflow.
| Evaluation | Muse Ember | Haiku 5.5 | Reported edge |
|---|---|---|---|
| AutomationBench-AA | 68.1% | 35.4% | Muse Ember |
| AA-Briefcase | 1,521 Elo | 1,574 Elo | Haiku 5.5 |
| Terminal-Bench 4.0 | 40.9% | 32.8% | Muse Ember |
| FrontierCode 2.0 | 27.2% | 34.9% | Haiku 5.5 |
Factual answers and refusals
On AA-Omniscience, which allows models to decline questions rather than guess, the results were mixed.
What drives the cost gap
Artificial Analysis attributes the lower index-task estimate to test usage, not simply to headline token rates.
The report lists both models at $0.10 per million input tokens and $0.50 per million output tokens.
Muse Ember’s cached input rate is listed at $0.004 per million tokens, 60% below Haiku’s.
In the index test, Ember used roughly a third as many output tokens as Haiku 5.5.
Lower Cost Comes With Trade-offs
For companies running high-volume, repetitive tasks, the reported price difference could make a lower-cost model worth testing. The benchmark’s $0.14 versus $0.41 figures suggest a substantially lower average cost for this particular index workload. But they do not establish savings for other tasks: actual costs depend on prompts, context length, caching, output volume and the quality bar a deployment needs to meet.
The measured strengths are uneven. On AutomationBench-AA, Muse Ember scored 68.1%, matching Meta’s more expensive Muse Wonder and exceeding Haiku 5.5’s 35.4% in the source’s comparison. On the GDP.pdf document test, Ember scored 22.8% against Haiku’s 20.8%, but it trailed Haiku on the AA-Briefcase knowledge-work test, with scores of 1,521 and 1,574 Elo respectively. Coding results also differed by test: Ember led on Terminal-Bench 4.0, 40.9% to 32.8%, but fell behind on FrontierCode 2.0, 27.2% to 34.9%.
That mix matters more than a single overall score for buyers choosing a model. A system that performs well on workflows but less well on factual recall may suit tasks grounded in supplied documents, while requiring a different model or human review for open-ended questions. Artificial Analysis also reports a Coding Agent Index score of 46 for Ember in Muse Code, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex. Those tool-specific results do not establish that Ember is better for every coding task.
Products Worth Considering
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
Apple Watch SE 3 GPS 40mm Smartwatch
As an affiliate, we earn on qualifying purchases.
How the Benchmark Prices Models
Muse Ember is described in the source as the efficient tier of Meta’s Muse model line, rather than Meta’s highest-scoring model. In the same Artificial Analysis comparison, Muse Wonder scored 44 and cost $6.95 per index task. The source also lists Claude Fable 5.2 at its highest effort level with a score of 65 and a task cost of $26.54. These comparisons show how benchmark cost can rise alongside score, but they do not establish a universal relationship between price and usefulness.
The benchmark’s task-cost estimate is distinct from the listed token price. An index task can involve repeated input, cached context and generated output, so models with similar per-token rates can have different total costs. Artificial Analysis says Muse Ember used less cached input cost and fewer output tokens than Haiku 5.5 during its tests. The reported savings therefore depend on the index’s prompts and usage pattern, not just Meta’s headline token rates.
Artificial Analysis’s AA-Omniscience test adds another dimension. It allows a model to decline questions rather than guess. Muse Ember answered 21.1% correctly and declined 64% of questions; its reported hallucination rate was 17.4%. Haiku 5.5 answered 34.7% correctly and had a 34.2% hallucination rate in the same comparison. These measures describe that test, not the probability of an error in every real-world interaction.
“The same benchmark work costs about a third.”
— Thorsten Meyer, ThorstenMeyerAI.com
Benchmark Results Have Limits
The figures cited here come from Artificial Analysis tests and model pages accessed on November 9 and 10, 2026, as reported by ThorstenMeyerAI.com. The source does not provide enough detail to independently reproduce the full evaluation or verify the cost estimates. Benchmark scores are not production guarantees, and it is not clear how closely the index’s task mix reflects any particular business’s workload.
The comparison also leaves practical questions open: how performance changes with different prompts, model settings, data sources and API usage; whether the listed prices remain current; and how often a deployment would need to route refused or uncertain answers to another system. The reported refusal and hallucination rates come from one factual-question test and should not be generalized to every subject or task.
Test Costs on Real Workflows
The immediate next step for prospective users is to evaluate Muse Ember on their own representative tasks, comparing cost, accuracy, refusal behavior, latency and the amount of human review required. For work based on company documents or retrieved information, testing should include whether the model uses supplied evidence reliably. For open-ended factual queries, teams should measure how refusals and incorrect answers affect the process.
Meta’s API and Muse Code are the listed routes for access in the source report. The report does not specify a later product milestone or a date for additional benchmark results. Current pricing, access terms and model behavior should be checked with Meta before deployment, and future independent evaluations may clarify how the benchmark results translate to everyday use.
Key Questions
What did Artificial Analysis report about Muse Ember’s score and cost?
For the max-effort setting, Artificial Analysis reported an Intelligence Index score of 36 and an average cost of $0.14 per index task. It reported the same score for Claude Haiku 5.5 at $0.41 per task. The cost figures are benchmark estimates, not fixed prices for every user workload.
Does Muse Ember have the same token prices as Claude Haiku 5.5?
The source report says both models list input at $0.10 per million tokens and output at $0.50 per million. It reports Muse Ember’s cached input price at $0.004 per million tokens, 60% less than Haiku’s, and says Ember generated fewer output tokens during the index test.
How does Muse Ember handle questions it cannot answer?
On Artificial Analysis’s AA-Omniscience test, Muse Ember declined 64% of questions. Its reported accuracy was 21.1% and hallucination rate 17.4% on that evaluation. These results apply to the test and do not predict every interaction.
Is Muse Ember better than Haiku 5.5 for coding?
The results are mixed. Muse Ember scored higher on Terminal-Bench 4.0, while Haiku 5.5 scored higher on FrontierCode 2.0. The source also reports a higher Coding Agent Index score for Ember in Meta’s Muse Code than for Haiku in Claude Code, but tool-specific scores do not show that one model will perform better on every coding job.
Source: ThorstenMeyerAI.com
