🔍 Read the full analysis: What Meta’s Muse Ember Gets Right—and Why It Often Says “I Don’t Know” on ThorstenMeyerAI.com
Get the latest gadgets 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 Nov. 9, 2026. Artificial Analysis measured its Intelligence Index score at 36, equal to Claude Haiku 5.5, with an average index-task cost of $0.14 versus $0.41. The benchmark also found low factual accuracy but a high refusal rate and a lower hallucination rate than several comparisons.
Meta released Muse Ember on Nov. 9, and an Artificial Analysis benchmark puts its Intelligence Index score at 36, level with Claude Haiku 5.5. The evaluator reports an average cost of $0.14 per index task for Muse Ember (max), compared with $0.41 for Haiku 5.5 (max); its tests also found that Ember often declines factual questions rather than answering them, limiting hallucinations but reducing how much it can answer.
Muse Ember is the efficient tier of Meta’s Muse model line. It is proprietary and available through Meta’s API and its Muse Code coding agent; it accepts text and images and returns text. Artificial Analysis tested it at two reasoning effort levels. At max, Ember scored 36 and cost $0.14 per index task; at xhigh, it scored 34 and cost $0.12. The evaluator says the task cost is the average API cost of completing one task in its index, not a general price for every customer workload.
The reported token prices are $0.10 per million input tokens and $0.50 per million output tokens, matching Haiku 5.5 and GPT-6 Luna in the source comparison. Cached input is listed at $0.004 per million tokens, which Artificial Analysis describes as 60% cheaper than Haiku’s. The evaluator attributes Ember’s lower index-task bill partly to cheaper cached input and partly to using about a third as many output tokens as Haiku to finish the index. It reports output speed of 440 tokens per second at max and a 1-million-token context window.
Results vary by task. Ember scored 68.1% on AutomationBench-AA, matching the score Artificial Analysis reports for Meta’s more expensive Muse Wonder; Haiku scored 35.4%. Ember led Haiku on Terminal-Bench 4.0, 40.9% to 32.8%, but trailed it on FrontierCode 2.0, 27.2% to 34.9%. In the evaluator’s coding-agent index, Ember within Muse Code scored 46, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex.
What Muse Ember Gets Right —and Why It Often Says “I Don’t Know”
Meta’s efficient-tier model matches Claude Haiku 5.5 on one composite benchmark at a lower average task cost. Its frequent refusals may limit hallucinations, but also reduce answer coverage.
Where Ember fits
The proprietary model is available through Meta’s API and Muse Code.
Lower benchmark bill
Ember’s average index-task cost was $0.14, compared with $0.41 for Claude Haiku 5.5.
A matched index score
At max reasoning effort, Ember and Haiku 5.5 both scored 36 on the Intelligence Index.
Text and image input
Ember accepts text and images, and returns text.
Performance depends on the task
Benchmark results show strengths in some workflows and weaker results in others.
| Evaluation | Ember | Haiku 5.5 |
|---|---|---|
| AutomationBench-AA | 68.1% | 35.4% |
| Terminal-Bench 4.0 | 40.9% | 32.8% |
| FrontierCode 2.0 | 27.2% | 34.9% |
| Coding-agent index | 46 in Muse Code | 37 in Claude Code |
The refusal trade-off
On AA-Omniscience, Artificial Analysis reports low factual accuracy alongside frequent abstention.
- A lower hallucination rate than several comparisons in the report.
- Refusals may help when a wrong answer could cause harm.
- Frequent refusals can leave factual questions unanswered.
- Teams may need a human reviewer or trusted fallback source.
Test it on your work
Benchmark averages are not a promise of performance or cost in a customer workflow.
Compare Ember with alternatives using your own documents and workflows.
Measure accuracy, refusals, latency and total cost—not just token prices or a composite score.
For factual work, test whether review or retrieval from a trusted source can handle unanswered questions.
Lower Costs, More Refusals
The result matters to teams weighing cost against answer coverage. If a system performs repetitive steps such as extraction, tagging, form filling or workflow automation, a lower benchmark-task cost could make it worth testing for high-volume use. But the index is not a measure of a company’s own workload, and the benchmark prices do not establish what a deployed application will cost after prompts, retries and human review.
Ember’s behavior on factual questions also changes the trade-off. Artificial Analysis reports 21.1% accuracy on its AA-Omniscience test and says the model declined 64% of questions, with a 17.4% hallucination rate. A refusal can be useful when a wrong answer would cause harm, provided the application can route unanswered questions to a person or another source. It can also interrupt a workflow that expects a complete response. These figures describe one benchmark, not a guarantee of how the model will behave in every setting.
Products Worth Considering
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
Bose Lifestyle Ultra Wireless Home Speaker, White Smoke
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
As an affiliate, we earn on qualifying purchases.
How Ember Compares
Artificial Analysis describes its Intelligence Index as an independent assessment based on nine evaluations. Its Nov. 9–10, 2026, model data places Ember (max) and Haiku 5.5 (max) at 36. The same comparison lists Meta’s Muse Wonder at 44 for $6.95 per index task, while Anthropic’s Claude Fable 5.2 at its highest effort scores 65 for $26.54. These figures make the lower-cost tier distinction clear, but scores from a composite index do not establish that one model is better for every task.
The factual-question results add another dimension. On AA-Omniscience, the source reports 34.7% accuracy and a 34.2% hallucination rate for Haiku 5.5; Gemini 3.8 Flash (high) scored 50.5% for accuracy. GPT-6 Luna’s reported hallucination rate was 79.3%. The test permits models to abstain, so accuracy, refusal frequency and hallucination rate need to be read together rather than treating accuracy alone as a complete measure of usefulness.
“For some jobs that is a weakness. For others it is exactly what you want.”
— Thorsten Meyer, source report
Limits of the Benchmark Evidence
The figures cited here come from Artificial Analysis evaluations and the source report, not from an independently verified account of customer deployments. The material does not provide confidence intervals, detailed test prompts or results across a broad range of real-world workloads. It is also unclear how frequently users will encounter refusals outside AA-Omniscience, or how the scores change with different prompts, tools and supplied information.
The reported costs are benchmark averages, and actual API bills will depend on usage patterns and implementation. The source material also begins to recommend fallback handling for refusals but ends before giving the full guidance. No customer adoption figures, independent business outcomes or detailed Meta response to the benchmark are included.
Testing Ember in Workflows
The next useful comparison is on the tasks an organization actually plans to run. Teams can test Ember against alternatives using their own documents and workflows, recording accuracy, refusal rates, latency and total cost, not just token prices or a composite benchmark score. For work that depends on correct factual answers, a review step or fallback to a trusted retrieval source may be needed when Ember abstains.
Further information could clarify how the model performs across customer deployments and whether Meta updates its pricing or benchmark results. Until then, the Nov. 9 release and Artificial Analysis results support a focused conclusion: Ember may offer lower-cost performance on some tasks, but its factual limitations and refusal behavior need to be tested before it is assigned work where a missing answer or a wrong one carries meaningful consequences.
Key Questions
What is Meta Muse Ember?
Muse Ember is the efficient tier in Meta’s proprietary Muse model line. It is available through Meta’s API and Muse Code, accepts text and images, and generates text.
How did Ember compare with Claude Haiku 5.5?
Artificial Analysis reported an Intelligence Index score of 36 for both models at max effort. Its average cost per index task was $0.14 for Ember and $0.41 for Haiku 5.5.
Why does Muse Ember often say it does not know?
On the AA-Omniscience test, Artificial Analysis reports that Ember declined 64% of questions. The test allows abstentions; the result does not establish how often Ember will refuse in other applications.
Does a lower benchmark cost mean Ember will be cheaper for every use?
No. The $0.14 figure is an average cost per task in Artificial Analysis’s index. An application’s actual costs depend on its prompts, context, output and workflow, so teams should measure their own usage.
Source: ThorstenMeyerAI.com
