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 November 9, 2026. Artificial Analysis measured a score of 36 at max effort and an average cost of $0.14 per index task, compared with $0.41 for Claude Haiku 5.5 at the same score. The benchmark also found low factual accuracy alongside a high refusal rate and lower hallucination rate; performance on individual business workloads remains unverified.
Meta released Muse Ember on November 9, and Artificial Analysis reports that its max-effort version scored 36 on the Intelligence Index at an average cost of $0.14 per index task. Claude Haiku 5.5 also scored 36 in the comparison, at $0.41 per task, making price and task performance the central point of the new model’s early benchmark results.
Muse Ember is the efficient tier of Meta’s proprietary Muse model line. It accepts text and images and produces text, and is available through Meta’s API and the Muse Code coding agent. Artificial Analysis tested it at two reasoning effort levels: Muse Ember max scored 36, while xhigh scored 34. The xhigh run cost $0.12 per index task and took 2.8 minutes, compared with $0.14 and 3.2 minutes for max. The source does not explain why the higher-effort setting scored lower.
The benchmark’s task cost reflects the average API cost of completing an index task, not a fixed price for every customer request. Artificial Analysis lists Muse Ember’s input and output prices at $0.10 and $0.50 per million tokens, respectively, matching the listed rates for Haiku 5.5 and GPT-6 Luna. Its cached input price is $0.004 per million tokens, which the source says is 60% below Haiku’s. Artificial Analysis also reports that Ember used roughly one-third as many output tokens as Haiku 5.5 across the index, helping explain the lower task cost despite matching token rates.
Results varied by task. Muse Ember scored 68.1% on AutomationBench-AA, matching Meta’s pricier Muse Wonder and exceeding Haiku 5.5’s reported 35.4%. It beat Haiku on Terminal-Bench 4.0, 40.9% to 32.8%, but trailed on FrontierCode 2.0, 27.2% to 34.9%. On GDP.pdf it scored 22.8% versus Haiku’s 20.8%, while its AA-Briefcase score was lower: 1,521 versus 1,574 Elo. These are benchmark results, not guarantees about customer deployments.
Meta’s Muse Ember Price and performance
Muse Ember matched Claude Haiku 5.5 on the Intelligence Index at a lower average task cost. Benchmark results vary by task and do not verify performance on business workloads.
Same score, lower task cost
Artificial Analysis reports matching token prices for the two models; lower cached input costs and fewer output tokens help explain Ember’s lower index task average.
- Score: 36
- Average task cost: $0.14
- About one-third as many output tokens as Haiku 5.5
- Score: 36
- Average task cost: $0.41
- Same listed input and output token rates
Results depend on the task
Ember led Haiku on some reported evaluations and trailed on others.
| Benchmark | Muse 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% |
| AA-Briefcase | 1,521 Elo | 1,574 Elo |
Check factual reliability
On AA-Omniscience, Ember’s factual results came with frequent declines.
Evaluate on real workflows
The index cost is a standardized benchmark average, not a quote for customer requests.
Include workflow, extraction and factual work your business actually needs.
Track accuracy, refusals, API usage, speed and human review.
Benchmark results do not guarantee performance on your workloads.
Lower Costs, Uneven Task Results
The figures suggest Muse Ember could be worth evaluating for repeated, narrowly defined tasks where a model processes information, follows a workflow or returns a structured result. At the reported index costs, Ember’s task average is about one-third of Haiku 5.5’s. That difference could matter for high-volume use, but actual costs depend on how a business uses the API, including input length, cached context and output volume.
The score does not establish that Ember is a general substitute for more capable models. Artificial Analysis reports a score of 44 for Muse Wonder at $6.95 per index task, while Claude Fable 5.2 scored 65 at its highest effort level for $26.54 per task. Those comparisons illustrate a price-performance range within this test; they do not establish that every additional point is worth the added expense or predict results on a particular company’s work.
Products Worth Considering
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
As an affiliate, we earn on qualifying purchases.
How the Index Compares Models
Artificial Analysis describes its Intelligence Index as a set of nine evaluations. Its task-cost figure is an average API cost for completing one task in that index. It is therefore a standardized comparison, not a quote for a user’s workload or a measure of total deployment cost. The available source reports results accessed November 9 and 10, 2026.
The comparison with Haiku 5.5 is notable because the two models share the same reported index score and token prices, while their average benchmark task costs differ. The source attributes much of the cost gap to cheaper cached input and lower output token use. It also reports that Muse Ember reached a score of 46 on the Artificial Analysis Coding Agent Index when used through Meta’s Muse Code, compared with 37 for Haiku 5.5 in Claude Code and 31 for GPT-6 Luna in Codex. Those coding-agent figures reflect specific model-and-tool pairings.
Accuracy and Real-World Fit
Artificial Analysis’s AA-Omniscience test found Muse Ember answered 21.1% of questions correctly and declined 64%. Its reported hallucination rate was 17.4%. In the same comparison, Haiku 5.5 answered 34.7% correctly and had a 34.2% hallucination rate; GPT-6 Luna’s reported hallucination rate was 79.3%. These are results from one benchmark, and the source does not provide enough detail here to establish how the measures translate to different kinds of factual work.
It remains unclear how Muse Ember will perform on customer workloads, how its API costs will vary with real usage, and whether the benchmark results will hold across other tests. The source says the model has a one-million-token context window and reports output speed of 440 tokens per second at max effort, but those figures do not establish how response quality or latency will vary in deployment. The source material also does not provide independent evaluations beyond the cited Artificial Analysis results.
Testing It on Business Work
Businesses considering Muse Ember should compare it with alternatives on representative tasks, tracking both accuracy and refusal rates as well as cost, speed and the amount of human review required. The benchmark’s results point to workflow and extraction tests as areas to examine, while its factual accuracy result calls for care when a task depends on obscure or specialized knowledge.
Further independent evaluations and customer results will show whether the reported price advantage carries over beyond the Artificial Analysis index. Meta’s release makes the model available through its API and Muse Code, but the source gives no timeline for additional benchmarks or changes to the model.
Key Questions
What did Meta announce?
Meta released Muse Ember, the efficient tier of its Muse model line, on November 9, 2026. It is available through Meta’s API and Muse Code.
How did Muse Ember compare with Claude Haiku 5.5?
Artificial Analysis reported that both models scored 36 on its Intelligence Index at max effort. Its average cost per index task was $0.14 for Ember and $0.41 for Haiku 5.5.
Does the $0.14 figure represent the price of every request?
No. It is the reported average API cost for a task in Artificial Analysis’s index. A customer’s cost will depend on actual usage, including token counts and cached input.
What does the benchmark say about factual answers?
On AA-Omniscience, Artificial Analysis reported that Muse Ember answered 21.1% correctly, declined 64% of questions and had a 17.4% hallucination rate. These results apply to that test and do not establish performance on every factual workload.
Source: Thorsten Meyer AI
