Meta's Muse Ember Scores 36 on the Intelligence Index at One-Third of Claude Haiku 5.5's Cost per Task

Meta's Muse Ember Scores 36 on the Intelligence Index at One-Third of Claude Haiku 5.5's Cost per Task

Meta's Muse Ember is the newest model in the company's efficient-tier line. At maximum reasoning effort, it scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 (max) while costing substantially less per task.

Muse Ember has the same standard input and output token prices as Claude Haiku 5.5 and GPT-6 Luna. Its cached-input pricing and lower output-token usage reduce its overall cost in the evaluated agentic workloads.

Key findings

• Strong position on the Intelligence Index and Cost per Task frontier: Muse Ember (max) is the least expensive model scoring at least 36, with a cost of $0.14 per task. At xhigh effort, Muse Ember scores 34 at $0.12 per task. Claude Haiku 5.5 (max) costs $0.41 for the same score, nearly three times as much. Muse Ember (max) is also approximately four times cheaper than DeepSeek V4.1 Flash 1017 (max), which scores 33 at $0.60, and GLM-5.3-Flash, which scores 28 at $0.58.
• Lower score than Muse Wonder at much lower cost: Muse Wonder (max) scores 44, eight points higher, but costs $6.95 per Intelligence Index task. Muse Ember's cost is approximately one-fiftieth as high.
• Strong AutomationBench-AA performance: Muse Ember (max) scores 68.1% on AutomationBench-AA, level with Muse Wonder (max). It also exceeds Claude Haiku 5.5 (max) on GDP.pdf, with scores of 22.8% and 20.8%, respectively. On AA-Briefcase, however, Muse Ember scores 1521 Elo, below Claude Haiku 5.5's 1574 Elo.
• Mixed agentic-coding results: Muse Ember exceeds Claude Haiku 5.5 on Terminal-Bench 4.0, scoring 40.9% compared with 32.8%, but trails it on FrontierCode 2.0, scoring 27.2% compared with 34.9%.
• Higher score in its dedicated coding agent: In Muse Code, Muse Ember (max) scores 46 on the Artificial Analysis Coding Agent Index. This is ahead of Claude Haiku 5.5 in Claude Code (max), at 37, and GPT-6 Luna in Codex (max), at 31.
• Knowledge remains its main weakness: Muse Ember records 21.1% accuracy on AA-Omniscience, the lowest result among the similarly intelligent models evaluated. Claude Haiku 5.5 (max) scores 34.7%, while Gemini 3.8 Flash (high) scores 50.5%. Muse Ember declines to answer 64% of the questions, which contributes to a hallucination rate of 17.4%, approximately half that of Claude Haiku 5.5 (max).

Pricing, reasoning, and availability

• Pricing: $0.10 per 1 million input tokens, $0.004 per 1 million cached input tokens, and $0.50 per 1 million output tokens
• Reasoning effort: Results are shown at maximum effort, with xhigh effort also included in the Intelligence Index versus Cost per Task analysis
• Availability: Meta's first-party API and Muse Code

Cost per task

Muse Ember (max) has the same input and output token prices as Claude Haiku 5.5 (max) and GPT-6 Luna (max). Its cached-input price is 60% lower, and cached input accounts for most of the cost in the agentic evaluations. Muse Ember (max) also produces approximately one-third as many output tokens as Claude Haiku 5.5 (max).

These factors result in a lower cost per task despite the same standard input and output token rates.

Automation and agentic coding

AutomationBench-AA is Muse Ember's strongest agentic result. Muse Ember (max) scores 68.1%, level with Muse Wonder and GPT-6 Astra.

Its agentic-coding performance varies by benchmark. Muse Ember (max) scores 40.9% on Terminal-Bench 4.0, ahead of Claude Haiku 5.5's 32.8%. On FrontierCode 2.0, it scores 27.2%, below Claude Haiku 5.5's 34.9%.

Performance as a coding agent

Within Muse Code, Muse Ember scores 46 on the Artificial Analysis Coding Agent Index. That result places it ahead of Claude Haiku 5.5 in Claude Code, which scores 37, and GPT-6 Luna in Codex, which scores 31.

The advantage comes from results on SWE-Atlas-QnA-AA and Terminal-Bench 4.0. Muse Ember remains behind Claude Haiku 5.5 on FrontierCode, consistent with their results on the broader Intelligence Index evaluations.

Knowledge and hallucination rate

Muse Ember's AA-Omniscience accuracy is 21.1%, the lowest among the similarly intelligent models evaluated. Its low accuracy is largely associated with declining to answer 64% of questions.

The same behavior keeps its hallucination rate at 17.4%, approximately half the rate recorded for Claude Haiku 5.5 (max). This produces a lower hallucination rate, but also reduces the model's answer coverage on knowledge-focused questions.

Evaluation summary

At maximum reasoning effort, Muse Ember combines an Intelligence Index score of 36 with a $0.14 cost per task. Its results are strongest in cost efficiency, AutomationBench-AA, and coding-agent evaluations through Muse Code. Its primary limitation in the reported tests is knowledge accuracy, while its refusal rate helps keep hallucinations relatively low.