🔍 Read the full analysis: Meta’s Muse Ember Delivers A 36 Score At 14 Cents Per Task 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 scores its max-effort version at 36 on its Intelligence Index, matching Claude Haiku 5.5, while estimating an average cost of $0.14 per index task versus $0.41 for Haiku. The test also found lower factual accuracy for Ember, alongside more frequent refusals and a lower hallucination rate.
Meta released Muse Ember on November 9, with the model scoring 36 on Artificial Analysis’s Intelligence Index and costing an estimated $0.14 per index task. That matches Claude Haiku 5.5’s score of 36, while Artificial Analysis lists Haiku’s average cost for the same benchmark work at $0.41; the comparison is based on the index, not a guarantee of costs or performance on customers’ own workloads.
Artificial Analysis tested Muse Ember at two reasoning-effort settings. The max setting scored 36 at $0.14 per index task, while xhigh scored 34 at $0.12. The index combines nine evaluations, and its task-cost figure is the average API cost of completing one task in that test suite. Artificial Analysis described max-effort Ember as the cheapest model it had found scoring 36 or higher on the index.
The model is Meta’s efficient tier in its Muse line, offered through Meta’s API and its Muse Code coding agent. It accepts text and images and generates text. Meta’s listed token 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 reports 440 output tokens per second and an average 3.2 minutes per index task for the max setting, as well as a one-million-token context window.
The price-per-task gap does not come from lower listed input and output rates: the source reports that those match Haiku 5.5. Artificial Analysis attributes Ember’s lower benchmark-task cost in part to cached input priced 60% below Haiku’s and to Ember using roughly one-third as many output tokens as Haiku to complete the index. Those figures describe performance and usage within the benchmark, not every application.
Meta’s Muse Ember The score and the trade-offs
Muse Ember matches Claude Haiku 5.5’s Intelligence Index score at a lower benchmark-task cost. Factual accuracy and refusal rates point to trade-offs that businesses should test on their own workloads.
A lower benchmark bill
On the same index, Ember and Haiku 5.5 both scored 36, while Ember’s reported average task cost was lower.
- Max effort: score 36 at $0.14 per task
- xhigh effort: score 34 at $0.12 per task
- Score 36 at $0.41 per task
- Listed token rates match Ember’s
Different strengths, different risks
Results vary by evaluation, so the overall score alone does not establish which model fits a workflow.
Accuracy and abstention
Ember answered 21.1% correctly and declined 64% of AA-Omniscience questions.
Lower reported rate
Its reported rate was 17.4%, versus 34.2% for Haiku 5.5 in that dataset.
Mixed performance
Ember matched Muse Wonder on AutomationBench-AA, beat Haiku on Terminal-Bench 4.0, and trailed it on FrontierCode 2.0.
Why the benchmark cost is lower
The reported gap reflects benchmark usage, not a guaranteed saving on customer workloads.
Artificial Analysis says Ember’s cached input is priced 60% below Haiku’s.
Ember used roughly one-third as many output tokens as Haiku to complete the index.
Caching, prompts, task difficulty and output length can change actual API costs.
Lower Costs, Different Trade-Offs
For teams that run high volumes of repetitive model calls, a lower cost on a comparable benchmark can make a new model worth testing for tasks such as routing, extraction, tagging and form filling. If an organization’s workload resembles the measured tasks, reducing per-task cost could matter at scale. But the benchmark does not establish that Ember will be three times cheaper on a company’s own data: token use, caching, prompts and task difficulty can change the bill.
The reported results point to a specific trade-off rather than an across-the-board replacement for more capable models. On Artificial Analysis’s AA-Omniscience factual test, 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 that dataset. A higher refusal rate may suit some workflows that can route unanswered questions to a person or another system, but it can also reduce usefulness when a direct answer is expected.
Task-level results are mixed. Ember scored 68.1% on AutomationBench-AA, matching the figure reported for Meta’s more expensive Muse Wonder and exceeding Haiku 5.5’s 35.4% in the source. It performed better than Haiku on Terminal-Bench 4.0, but worse on FrontierCode 2.0. The split makes evaluation on the intended task more informative than relying on the overall index score alone.
Products Worth Considering
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
DJI Osmo Mobile 7P Gimbal Stabilizer with Mic Mini Transmitter, Infinity Black
As an affiliate, we earn on qualifying purchases.
How Ember Compares on the Index
Artificial Analysis’s figures place Muse Ember (max) alongside models with the same overall index score but different measured task costs. Its comparison lists Claude Haiku 5.5 at 36 and $0.41 per task, DeepSeek V4.1 Flash 1017 at 33 and $0.60, and GPT-6 Luna at 28 and $0.15. Meta’s larger Muse Wonder scores 44 at $6.95 per task. These are the source’s reported model-page and benchmark figures, accessed November 9 and 10, 2026.
The gap widens higher up the index, though the scores and task costs are not a direct measure of value for every user. Artificial Analysis lists Muse Wonder as eight points above Ember at roughly 50 times its per-task cost. It lists Claude Fable 5.2 at its highest effort level with a score of 65 and a cost of $26.54 per index task. These comparisons show the price attached to higher scores in this particular dataset; they do not establish that the more expensive models are better choices for each application.
The component tests add detail beyond the composite score. Ember scored 22.8% on GDP.pdf, compared with Haiku’s 20.8%, but trailed Haiku on the AA-Briefcase knowledge-work test in Elo rating, 1,521 to 1,574. In the Artificial Analysis Coding Agent Index, the source reports 46 for Ember inside Muse Code, 37 for Haiku inside Claude Code and 31 for GPT-6 Luna inside Codex. The source does not provide enough methodological detail here to treat those results as a general ranking across coding tasks.
What the Benchmarks Do Not Show
The supplied source reports results from Artificial Analysis, but does not include independent replication, full benchmark protocols or confidence intervals for the figures. The scores should be read as benchmark results reported by that evaluator, not proof that the models will tie on every real-world task. The source also does not provide a broader set of customer deployments or independently audited operating costs.
It remains unclear how often the 64% refusal rate on AA-Omniscience will carry over to other factual questions or business workflows. A refusal may prevent a fabricated answer, but the available figures do not show whether Ember can distinguish reliably between questions it cannot answer and ones it could answer with supplied material. The results also do not establish how often the model will abstain when connected to retrieval tools or a company database.
Actual costs depend on usage patterns, including how much context is cached, how many output tokens a task requires and which API settings are used. The source provides listed token prices and benchmark averages, but not a complete cost comparison across customer workloads. It also does not specify the full availability terms or whether Meta plans changes to pricing or access.
Testing Ember in Real Workflows
The next practical step for prospective users is to test Muse Ember on representative tasks and compare it with alternatives using the same prompts, inputs and success criteria. Teams can track accuracy, refusal rates, latency and total API cost rather than assuming that the index’s $0.14 average will apply to their own workloads.
For deployments where an incorrect answer carries risk, evaluations should include both ordinary cases and questions the model should decline. A workflow might provide Ember with relevant documents or retrieved records, then send unresolved cases to another model or a human reviewer. That approach could take advantage of its reported low cost and refusal behavior, but its reliability would need to be measured in the target setting.
Meta’s release makes Muse Ember available through the company’s API and Muse Code, according to the source. Further evidence will depend on customer testing, additional evaluator results and any updates to access or pricing. The source does not announce a later benchmark release or a specific next product milestone.
Key Questions
What did Meta announce?
Meta released Muse Ember on November 9, 2026. The model is described as the efficient tier of its Muse line and is available through Meta’s API and Muse Code, according to the source.
What does Muse Ember’s score of 36 mean?
It is Muse Ember’s score at max reasoning effort on the Artificial Analysis Intelligence Index, which the source says covers nine evaluations. Claude Haiku 5.5 also scored 36 in the reported comparison.
Why is Ember’s reported task cost lower than Haiku’s?
Artificial Analysis attributes the difference to cheaper cached input and lower output-token use during the index. The listed input and output token prices are the same in the comparison; the $0.14 and $0.41 figures are average costs per index task, not universal prices for a customer request.
Does Muse Ember hallucinate less than Claude Haiku 5.5?
In the reported AA-Omniscience results, Ember’s hallucination rate was 17.4%, compared with 34.2% for Haiku 5.5. Ember also answered fewer questions correctly and declined more often, so the figures do not mean it is more accurate overall.
Is Muse Ember a better choice for coding?
The reported coding results are mixed. Ember scored higher than Haiku on Terminal-Bench 4.0 but lower on FrontierCode 2.0. The source also reports a higher Coding Agent Index result for Ember in Muse Code than for Haiku in Claude Code, but that does not establish superiority across coding work generally.
Source: ThorstenMeyerAI.com
