🔍 Read the full analysis: Muse Ember At 14 Cents Per Task: A 36 Score With An Important Caveat on ThorstenMeyerAI.com
Get smart everyday buys 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 setting at 36, matching Claude Haiku 5.5, while estimating an index-task cost of $0.14 versus $0.41 for Haiku. The benchmark also found low factual accuracy but a high rate of declining to answer, so the price and score do not establish that it will suit every workload.
Meta released Muse Ember on November 9, and Artificial Analysis rated its max reasoning setting at 36 on the Intelligence Index, the same score as Claude Haiku 5.5 (max). The evaluator estimates an average cost of $0.14 per index task for Ember, compared with $0.41 for Haiku, but also reports substantially lower factual accuracy on a separate knowledge test.
Artificial Analysis says its Intelligence Index combines nine evaluations. In the comparison reported on November 9 and 10, Muse Ember (max) scored 36, while the xhigh setting scored 34. The model accepts text and images and produces text; Meta offers it through its API and its Muse Code coding agent. These results are benchmark findings, not a guarantee of performance or cost on a particular customer’s tasks.
The reported cost per index task is an average API cost for completing one task in the evaluator’s index. Artificial Analysis attributes Ember’s lower total to cached input priced at $0.004 and to using roughly a third as many output tokens as Haiku 5.5 during the index. The source lists Ember’s standard input and output rates as $0.10 and $0.50 per million tokens, respectively, the same as Haiku 5.5 and GPT-6 Luna. It reports output speed of 440 tokens per second, a 3.2-minute average per index task, and a one-million-token context window for Ember.
Results vary by task. Artificial Analysis reports 68.1% on AutomationBench-AA, matching Meta’s more expensive Muse Wonder and exceeding Haiku 5.5’s 35.4%. Ember scored higher than Haiku on Terminal-Bench 4.0, but lower on FrontierCode 2.0. On the evaluator’s Coding Agent Index, Ember in Muse Code scored 46, compared with Haiku in Claude Code at 37 and GPT-6 Luna in Codex at 31. These are results from different tests and product configurations, so they should not be treated as a single, universal ranking.
Muse Ember’s benchmark case low cost, important caveat
Meta’s model matched Claude Haiku 5.5 on Artificial Analysis’s Intelligence Index at a lower estimated task cost. Separate tests raise questions about factual accuracy and unanswered queries.
The cost comparison
Artificial Analysis estimates average cost for its index tasks, not for every customer workload.
Performance depends on the task
The benchmarks show strengths in some workflows, not a universal ranking.
AutomationBench-AA
Ember scored 68.1%, matching Muse Wonder and exceeding Haiku 5.5’s 35.4%.
Mixed results
Ember scored higher than Haiku on Terminal-Bench 4.0, but lower on FrontierCode 2.0.
Muse Code scored 46
The comparison reported Haiku in Claude Code at 37 and GPT-6 Luna in Codex at 31.
The factuality caveat
On AA-Omniscience, the evaluator reported both frequent abstentions and limited correct answers.
| Measure | Muse Ember | Claude Haiku 5.5 |
|---|---|---|
| Correct answers | 21.1% | Not stated |
| Declined questions | 64% | Not stated |
| Reported hallucination rate | 17.4% | 34.2% |
What teams should test
Evaluate Ember on representative tasks before routing work to it.
Test representative prompts and provide relevant facts through documents, retrieval, or databases.
Measure accuracy, abstentions, latency, cache use, and total API cost.
Assess human review or another model for queries Ember leaves unresolved.
Lower Benchmark Costs, Different Risks
The headline comparison may matter to teams running high-volume, repeatable AI tasks: Artificial Analysis’s estimated index-task cost for Ember is about one-third of Haiku 5.5’s, despite the two models receiving the same overall score. If that pattern carries over to a company’s own workflow, lower costs could make frequent classification, extraction, or agent steps more affordable. The evaluator’s price is specific to its index, however; actual costs depend on prompts, context length, caching, output, and the way a system is configured.
The knowledge-test results point to a different deployment consideration. On AA-Omniscience, Artificial Analysis reports Ember answered 21.1% correctly and declined 64% of questions. Its reported hallucination rate was 17.4%, compared with 34.2% for Haiku 5.5. A system that often abstains may be useful where unsupported answers are costly, but it can also leave more queries unresolved and require a human or another model to respond. The figures describe one evaluator’s test, not the rate users should expect in every setting.
That makes the practical choice less about a single score than about the job. Ember’s reported workflow results and lower index cost may appeal where the model receives the relevant facts in a document or retrieved context. Its comparatively weak performance on general factual questions cautions against relying on it as a stand-alone source of stored knowledge. Businesses should test their own prompts and track both task success and abstentions, as well as cost.
Products Worth Considering
Apple iPad 11-inch (A16), 128GB, Wi-Fi, Silver
Dell UltraSharp U2724DE 27-Inch WQHD IPS Black Monitor
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
As an affiliate, we earn on qualifying purchases.
How the Scores Compare
Muse Ember is described by its source as the efficient tier of Meta’s Muse model line and a proprietary model available through Meta’s API and Muse Code. Artificial Analysis tested it at max and xhigh reasoning effort. The max setting scored 36 at $0.14 per index task; xhigh scored 34 at $0.12. The source gives output speeds of 440 and 423 tokens per second for the two settings, respectively, and average task times of 3.2 and 2.8 minutes. These paired results show that the lower-cost setting also scored lower on the index; they do not establish which option is best for other tasks.
For comparison, Artificial Analysis lists Muse Wonder (max) at 44 points and $6.95 per task, while Claude Fable 5.2 at its highest effort scored 65 at $26.54 per task. The source characterizes the latter cost as about 190 times Ember’s for 29 additional points. Those numbers compare specific model settings under the index methodology, not the full cost or value of deploying each system. Ember and Haiku’s matching index scores likewise do not mean their capabilities are interchangeable: the evaluator found meaningful differences across workflow, document, coding, and knowledge tests.
Benchmark Limits and Open Questions
The reported figures come from Artificial Analysis’s benchmarks and model pages; the supplied material does not provide independent replication or results from a customer’s production workload. It is not clear how closely the index’s prompts, context, caching, and task mix match any particular use case. The per-task cost is an evaluator average, not a fixed price for every request.
The source also does not give enough detail to establish how the reported scores translate to real-world reliability across industries, languages, or changing model versions. A high refusal rate can reduce unsupported answers, but the supplied results do not quantify how often a refusal is appropriate, how frequently a system would need a fallback, or what that would add to total cost. Those questions remain for deployment testing.
Test Workloads Before Routing
The immediate next step for teams considering Ember is to compare it with alternatives on their own representative tasks, rather than treating the Intelligence Index as a substitute for evaluation. Testing should include accuracy, refusal rates, latency, cache use, total API cost, and the performance of any fallback model or human review process. The source advises measuring against a company’s own workload and supplying relevant information through documents, retrieval, or databases when the task depends on facts.
The supplied material does not announce a further benchmark release, pricing change, or scheduled Meta update. Meta’s API and Muse Code are the stated access routes. Further evidence about practical performance will depend on user testing and any later evaluations that report methods and comparable results.
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, which uses nine evaluations. The xhigh setting scored 34.
How does its reported task cost compare with Claude Haiku 5.5?
Artificial Analysis estimates $0.14 per index task for Muse Ember (max) and $0.41 for Claude Haiku 5.5 (max). Those are benchmark-specific average costs, not guaranteed charges for other workloads.
What is the main caveat in the benchmark results?
On AA-Omniscience, Artificial Analysis reports that Ember answered 21.1% correctly and declined 64% of questions. Its reported hallucination rate was 17.4%; the supplied results do not show how these rates will translate to every use case.
Where did Muse Ember perform well?
Artificial Analysis reports 68.1% on AutomationBench-AA, level with Muse Wonder, and a score of 46 for Ember in Meta’s Muse Code on its Coding Agent Index. Results were mixed across other document and coding tests.
Should businesses switch to Muse Ember based on these results?
The benchmark alone cannot establish that. Teams should test Ember on their own tasks and account for accuracy, refusals, latency, caching, and fallback costs before making a deployment decision.
Source: ThorstenMeyerAI.com
