🔍 Read the full analysis: Meta’s Muse Ember Offers A 36 Score, But Its Answers May Be Uncertain 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 November 9, 2026. Artificial Analysis measured its Intelligence Index score at 36, matching Claude Haiku 5.5, with an average index-task cost of $0.14 versus $0.41. The evaluation also found low factual accuracy alongside frequent refusals and a lower hallucination rate than several compared models.
Meta released Muse Ember on November 9, and an Artificial Analysis evaluation gave its max-reasoning version an Intelligence Index score of 36, the same score as Claude Haiku 5.5. Artificial Analysis reported an average cost of $0.14 per index task for Ember, compared with $0.41 for Haiku, while also finding that Ember answered few questions correctly on a factual-knowledge test and often declined to answer.
Artificial Analysis says the Intelligence Index combines nine evaluations. Its reported task cost is the average API cost of completing one task in that index, not a guarantee of the cost for a business’s particular workload. The comparison is based on the models and settings tested on November 9 and 10, 2026. Muse Ember is a proprietary Meta model available through Meta’s API and in its Muse Code coding agent. It accepts text and images and produces text.
At max reasoning effort, Ember scored 36 and cost $0.14 per index task. At xhigh effort, it scored 34 at $0.12 per task. Artificial Analysis reported 440 output tokens per second and 3.2 minutes per index task for the max setting on Meta’s API, along with a one-million-token context window. Listed prices are $0.10 per million input tokens and $0.50 per million output tokens; cached input is listed at $0.004 per million tokens.
Artificial Analysis attributes Ember’s lower cost per index task, despite token prices matching Haiku 5.5’s, partly to cheaper cached input and lower output-token use during the index. It says Ember used roughly a third as many output tokens as Haiku 5.5 to complete the benchmark. These are evaluator-reported results, not a direct measure of cost or performance on every customer’s work.
Muse Ember scores 36 but answers may be uncertain
Meta’s model matched Claude Haiku 5.5 on Artificial Analysis’s Intelligence Index at a lower reported task cost. Its factual test also found frequent refusals and low accuracy.
What the headline comparison means
The index combines nine evaluations; its task cost is not a quote for every business workload.
- Score: 36 at max reasoning
- Reported index-task cost: $0.14
- Lower hallucination rate than Haiku on the cited knowledge test
- Same index score: 36
- Reported index-task cost: $0.41
- Higher factual accuracy and fewer refusals on that test
A mixed factual-answer profile
On a test that allows abstentions, Ember declined many questions while recording a lower hallucination rate than several compared models.
Performance varies by task
The shared index score does not establish which model will work best for a company’s specific tasks.
Bounded, repeated work
The lower benchmark-task cost may justify testing extraction, tagging, and workflow steps.
Frequent abstentions
Users may need to provide evidence, route requests elsewhere, or add human review.
Test task by task
Ember performed differently across coding, automation, and knowledge-work evaluations.
Test before relying on the headline
Benchmark results may not transfer to different prompts, data, or production workloads.
Compare Ember with current systems using the same prompts and task sets.
Track quality, cost, abstentions, escalation, latency, and human review.
Check how it uses supplied documents and decide when to route a request to another model or a person.
Lower Cost, Different Trade-Offs
The headline comparison gives organizations a reason to test Ember for repeated, relatively bounded tasks, such as extraction, tagging, and workflow steps. In Artificial Analysis’s reported figures, its index-task cost is roughly one-third of Haiku 5.5’s, while the models share a score of 36. That could matter in high-volume use, but actual savings depend on prompts, context length, caching, retries, and the model’s success rate on a company’s own tasks.
The benchmark also points to a trade-off between stored factual knowledge and willingness to answer. On AA-Omniscience, a knowledge test that allows models to abstain, 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 in the same data. A refusal may reduce unsupported answers, but it can also mean a user must supply evidence, use retrieval, or route the request to another system. The figures should not be treated as a universal safety guarantee.
Results were not uniformly favorable across tasks. Artificial Analysis reported 68.1% on AutomationBench-AA, level with Meta’s more expensive Muse Wonder, and stronger performance than Haiku 5.5 on Terminal-Bench 4.0. Ember scored lower than Haiku on FrontierCode 2.0 and on the AA-Briefcase knowledge-work test. The mixed profile makes task-specific testing more relevant than the single index score.
Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

Samsung Galaxy Tab A11+ Kids Edition, 128GB, Royal Blue

Samsung 32 Odyssey G5 QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Ember Compares on Tests
Artificial Analysis’s reported comparison places Muse Ember (max) alongside models with similar scores or prices. Claude Haiku 5.5 (max) also scored 36, at $0.41 per index task. GPT-6 Luna (max) scored 28 at $0.15 per task, while Meta’s Muse Wonder (max) scored 44 at $6.95. The source also reports that Claude Fable 5.2 at its highest effort scored 65 at $26.54 per task. These cost figures refer to the index evaluation, not identical real-world workloads.
On the factual-knowledge test, Artificial Analysis reported accuracy of 34.7% for Haiku 5.5 and 50.5% for Gemini 3.8 Flash at high effort, compared with Ember’s 21.1%. The evaluator reported Ember’s 17.4% hallucination rate, Haiku’s 34.2%, and GPT-6 Luna’s 79.3% on the same test. Because the test permits abstentions, accuracy and hallucination figures need to be read alongside Ember’s 64% refusal rate, not as standalone measures of general reliability.
In coding-agent results, the source says Ember scored 46 on Artificial Analysis’s Coding Agent Index when used in Meta’s Muse Code, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex. Those figures compare models within different named coding tools, so they do not establish that Ember will outperform them in every coding setup.
Limits of the Benchmark Results
The figures available here come from Artificial Analysis’s evaluations; the source material does not include an independent replication or Meta’s response to the results. It is not clear how Ember performs across a wider range of languages, domains, prompts, or production workloads, or whether the reported cost advantage holds when usage patterns differ from the index.
The benchmark summary also does not establish why Ember declines so often, how users can adjust its refusal behavior, or how consistently the lower hallucination rate would carry over to real applications. Refusal and hallucination rates describe results under a specific test design. They do not prove that a model will abstain appropriately in every setting or that an answer it does provide is correct.
More broadly, comparisons involving different tools and test suites have limits. The reported Coding Agent Index results involve Muse Code, Claude Code, and Codex, while the workload and configuration details needed to generalize from those scores are not provided in the supplied material. Customer-side testing remains necessary before relying on the model or estimating savings.
Testing Ember in Real Workflows
Meta’s model is available through its API and Muse Code, according to the source. The next practical step for organizations considering it is to test representative tasks against their current systems, measuring not only answer quality and cost but also abstentions, escalation rates, latency, and human review. Comparisons should use the same prompts and task sets, with clear rules for when a refusal counts as an acceptable result.
For work that depends on supplied documents or company data, teams can assess Ember with that information provided directly or through retrieval, then check whether it extracts and uses the evidence correctly. A fallback to a different model or human review may be needed when Ember declines or when the task requires knowledge absent from the supplied material. Further benchmark updates or published deployment results could clarify how well the reported price and performance profile transfers beyond Artificial Analysis’s tests; no such next release or evaluation date is specified in the available source.
Key Questions
What is Meta Muse Ember?
Muse Ember is a proprietary Meta model released November 9, 2026. The source says it is available through Meta’s API and Muse Code, accepts text and images, and generates text.
What score did Muse Ember receive?
Artificial Analysis reported an Intelligence Index score of 36 for Muse Ember at max reasoning effort. Its xhigh setting scored 34. The index combines nine evaluations, according to the source.
How much does an index task cost?
Artificial Analysis reported an average of $0.14 per index task for Muse Ember at max effort and $0.41 for Claude Haiku 5.5 at max. These are benchmark-task cost estimates, not a quote for every real-world use.
Does Muse Ember hallucinate less than Haiku 5.5?
In Artificial Analysis’s AA-Omniscience results, Ember’s reported hallucination rate was 17.4%, compared with 34.2% for Haiku 5.5. Ember also answered only 21.1% correctly and declined 64% of questions in that test, so the figures do not establish its reliability across all tasks.
Is Muse Ember better for coding?
The reported results are mixed. Ember scored higher than Haiku 5.5 on Terminal-Bench 4.0 but lower on FrontierCode 2.0. The source also reports a higher Coding Agent Index score for Ember in Muse Code than for Haiku in Claude Code, a comparison that should not be generalized to every coding task or setup.
Source: ThorstenMeyerAI.com
