Meta Superintelligence Labs has released Muse Ember, a hosted multimodal reasoning model designed for the repetitive steps inside agents. The tool targets tool calls, coding subtasks, routing, extraction and long document parsing. It scores 36 on the Artificial Analysis Intelligence Index, level with Claude Haiku 5.5 at max effort. Meta puts its cost at about $0.14 per index task.
TL;DR
- Size: 1M-token context window, up to 512K output tokens.
- Runs on: Meta’s hosted infrastructure only (Meta Model API and Muse Code). Weights are private, so there is no local GPU, laptop or quantized option.
- Performance: Ties Claude Haiku 5.5 (max) at 36 on the AA Intelligence Index, ahead of GPT-6 Luna (28).
- Best: AutomationBench-AA at 37.8, versus 20.2 for GPT-6 Luna and 13.9 for Claude Haiku 5.5.
- Worst: FrontierCode 2.0-AA at 27.2, behind Claude Haiku 5.5 at 34.9.
- Bottom line: Best: Haiku-level intelligence at $0.136 per task versus $0.411. Worst: closed weights and weaker on mergeable real-repo code.
What is Meta Muse Ember?
Muse Ember is a small reasoning model from Meta for high-volume agent subtasks. It is the fourth model under the Muse brand, after Spark, Glimmer and Wonder.
Meta positions it below its larger models. Muse Wonder handles the hardest problems and Muse Spark covers coding and agentic workflows. Ember takes repeatable coding and agent tasks at scale. Meta’s own framing: Muse Wonder plans, Muse Ember handles the volume.
How does Muse Ember work?
Muse Ember is the first model Meta trained using Muse Wonder Full, its largest model.
The model natively accepts text, images, audio and video. That lets an agent read screenshots, charts, scanned pages, voice memos and meeting recordings without a separate model. It also operates desktop apps from screenshots.
The 1M-token context window fits a full codebase or document set in one call. Output runs up to 512K tokens, 4x the 128K cap on both rivals.
How does Muse Ember perform on benchmarks?
All results below are at max reasoning effort, as reported by Meta. Some come from Artificial Analysis (AA), others from Meta’s internal suite.
- Agents: On AutomationBench-AA, 657 Zapier business workflows, Ember completes 37.8% of tasks. GPT-6 Luna reaches 20.2% and Claude Haiku 5.5 reaches 13.9%. On Agents’ Last Exam (internal), Ember scores 61.6 versus 57.1 and 57.0.
- Coding: Ember leads Terminal-Bench 4.0 with 40.9 versus Haiku’s 39.2 and Luna’s 16.4. It also leads the AA Coding Agent Index, 46 to 37 to 31. It loses FrontierCode 2.0-AA, where Haiku scores 34.9 and Ember 27.2.
- Multimodal: Ember ties GPT-6 Luna on GDP.pdf-AA at 22.8. It edges Haiku on OSWorld 2.1, 69.5 versus 68.9 on partial credit. PerceptionBench with tools gives Ember 80.6, Haiku 78.6 and Luna 62.8.
Why is Muse Ember cheaper per task if token prices match?
Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens. Cached input costs $0.004 per million. Claude Haiku 5.5 and GPT-6 Luna list the same base rates.
With identical list prices, the cost gap must come from tokens used per task. On Meta’s chart, Ember at max scores 35.8 for $0.136 per task. Haiku 5.5 at max scores 36.2 for $0.411. Luna’s best setting scores 27.8 for $0.152. Artificial Analysis summarised this as matching Haiku at a third of the cost per task.
Long prompts widen the gap on paper. Haiku 5.5 charges $0.50 and $2.50 above 100K prompt tokens. GPT-6 Luna doubles input rates above 272K input tokens. Meta lists no long-context tier for Ember.
Muse Ember vs Claude Haiku 5.5 vs GPT-6 Luna
| Feature | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| Developer | Meta | Anthropic | OpenAI |
| Released | Nov 9, 2026 | Oct 7, 2026 | Sep 22, 2026 |
| Parameters | Not disclosed | Not disclosed | Not disclosed |
| Context window | 1M | 1M | 1.05M |
| Max output | 512K | 128K | 128K |
| Inputs | Text, image, audio, video | Text, image | Text, image |
| Weights | Closed | Closed | Closed |
| Input / output per 1M | $0.10 / $0.50 | $0.10 / $0.50 (up to 100K prompt) | $0.10 / $0.50 (up to 272K input) |
| Cached input per 1M | $0.004 | $0.01 | $0.01 |
| AA Intelligence Index (max)* | 36 | 36 | 28 |
| Cost per AA task (max)* | $0.136 | $0.411 | $0.152 |
| AutomationBench-AA* | 37.8 | 13.9 | 20.2 |
| Terminal-Bench 4.0* | 40.9 | 39.2 | 16.4 |
| FrontierCode 2.0-AA* | 27.2 | 34.9 | 25.2 |
| OSWorld 2.1 (partial / binary)* | 69.5 / 33.8 | 68.9 / 27.8 | 50.5 / 17.0 |
*Benchmark rows as published by Meta in its launch post and evaluation report, sourced from Artificial Analysis Intelligence Index v5 and Meta internal runs.
How can developers access Muse Ember?
Ember is available on the Meta Model API and in Muse Code, Meta’s coding agent. It supports the Responses API. Partners in early testing included Box, Cline, Oracle, Cloudflare and You.com.
What it means
For people building agents, the model offers the same intelligence as the current market leader for automation tasks but at a fraction of the cost. The ability to ingest video and audio directly means developers can skip separate transcription steps for meeting notes or voice memos. The 1M-token window allows a single prompt to handle a full repository or large document set, reducing the overhead of chunking data. However, the lack of local deployment options means teams must rely on Meta’s hosted infrastructure. Additionally, the model underperforms on complex, mergeable code tasks compared to its peers, suggesting it is best suited for execution rather than deep architectural changes.
