Meta releases Muse Ember for high-volume agent tasks

A model for high-volume tasks

Meta Superintelligence Labs has introduced Muse Ember, a model specifically designed to handle the repetitive, high-volume steps required by AI agents. While larger models like Muse Wonder are built for complex planning, Ember is intended to manage the heavy lifting of subtasks. These include things like routing, data extraction, and parsing long documents.

The model is the fourth entry in the Muse brand, following Spark, Glimmer, and Wonder. Meta positions it as a smaller and more efficient option within their ecosystem. It aims to provide a balance between intelligence and scale for developers who need to run many small tasks quickly.

By focusing on these specific workflows, Meta hopes to make agentic behavior more practical for businesses. It is built to handle coding subtasks and tool calls that often require high reliability but do not necessarily need the full reasoning power of a flagship model.

Multimodal capabilities and scale

Muse Ember is a multimodal reasoning model, meaning it can natively process text, images, audio, and video. This allows an AI agent to interact with a variety of inputs, such as reading screenshots, analyzing charts, or listening to voice memos without needing a separate model for each format. It can even operate desktop applications by looking at screenshots.

One of its most significant technical specs is the 1M-token context window. This allows the model to ingest a full codebase or a massive set of documents in a single request. Additionally, it supports an output of up to 512K tokens, which is four times the limit found on its primary rivals.

The model was trained using Muse Wonder Full, Meta's largest model. This lineage helps it maintain high performance across different benchmarks. Despite its efficiency, it remains a closed-weights model, meaning it can only be accessed through Meta's hosted infrastructure and does not have a local GPU or quantized version.

Benchmarking performance

In terms of general intelligence, Muse Ember scores 36 on the Artificial Analysis Intelligence Index. This puts it on par with Claude Haiku 5.5 at maximum effort and ahead of GPT-6 Luna, which scored 28. It shows particular strength in automation, completing 37.8% of tasks on the AutomationBench-AA, significantly outperforming its competitors in that category.

The model also shows promise in coding tasks, leading the Terminal-Bench 4.0 with a score of 40.9. However, it does face some limitations in specific areas. For instance, it scored 27.2 on the FrontierCode 2.0-AA benchmark, trailing behind Claude Haiku 5.5, which scored 34.9 in that specific category of mergeable real-repo code.

Meta's internal tests also show strong results. On the Agents’ Last Exam, Ember scored 61.6, slightly beating both Claude Haiku 5.5 and GPT-6 Luna. These results suggest that while it may not be the best at every single niche, it is highly competitive for the specific agentic workflows Meta is targeting.

Cost efficiency for developers

A major highlight for developers is the cost-to-performance ratio. While Muse Ember shares the same base token prices as its rivals—$0.10 per million input tokens and $0.50 per million output tokens—it is significantly cheaper per task. As marktechpost.com reported, the model costs about $0.136 per index task, which is roughly one-third of the $0.411 cost for Claude Haiku 5.5.

This price gap exists because Muse Ember is optimized to use fewer tokens to complete a task. Long prompts can often drive up costs on other models, such as Haiku 5.5, which charges extra for prompts over 100K tokens. Meta does not list a long-context tier for Ember, which may help keep costs predictable for large-scale operations.

For companies running high-volume agentic workflows, this efficiency is a key differentiator. It allows for more frequent calls and larger-scale automation without the costs scaling linearly with the complexity of the prompts. It makes the