Muse Ember: Meta’s small model designed for repetitive tasks
Meta introduced Muse Ember, its smallest model, designed for repetitive tasks of AI agents. Available in Muse Code and via the Meta API, it focuses on cost, speed, and multimodal processing.
Meta introduced Muse Ember on November 9, 2026, as its smallest model and most efficient option for repetitive tasks of AI agents. Available the same day in Muse Code and via the Meta Model API, it targets tool calls, code subtasks, extraction, and analysis of long documents. The goal is simple: have the model perform mechanical actions while heavier systems keep the brain of the operation.
A model designed for execution, not for grand speeches
Muse Ember, presented under the name muse-ember-1.0, is designed for the work layer that occurs most often in agents: reading, sorting, routing, correcting, retrying. Meta says its agents spend most of their tokens on these short and structured sequences. Ember can also chain tools, retrieve records, and update systems across multiple applications, making it useful for executing a business flow from start to finish.
The model supports text, image, audio, and video. In practice, it can therefore process screenshots, tables, scans, voice notes, or meeting recordings without calling on another model. This versatility matters because it broadens its field of action to the workstation, not just code.
Pricing, context, and division of labor at Meta
On paper, Meta offers aggressive pricing: $0.10 per million input tokens, $0.004 for cached tokens, and $0.50 per million output tokens. Ember accepts a context of 1 million tokens, enough to encompass a complete codebase or a document set in a single request, and can produce up to 512,000 output tokens.
Meta also contrasts it with Muse Spark 1.3, with a fairly clear specialization logic. Wonder is meant to plan the hardest problems, Spark handles code and agent tasks, while Ember executes large-scale repetitive versions of those same tasks. The promise is coherent. The classic small catch remains: a very efficient model does not magically become better on all subjects, as its real code results remind us.
Solid scores, but not across the board
In comparisons published by Meta, Ember scores 36 on the Artificial Analysis Intelligence Index, versus 37 for Muse Spark 1.3, with an estimated cost per task of $0.14 versus $2.39 for Spark. The model also presents itself as more economical than Claude Haiku 5.5 and GPT-6 Luna at the same reasoning level, according to Meta’s estimates based on its public prices. However, FrontierCode 2.0-AA places it behind Claude Haiku 5.5 on real code tasks, with 27.2% versus 34.9%.
Other measures go in the same direction. Meta announces 37.8% on AutomationBench-AA for business workflows, 61.6% on Agents’ Last Exam in the Linux subset, 40.9% on Terminal-Bench 4.0, and 46 on the Coding Agent Index-AA. On OSWorld 2.1, Ember scores 69.5% in partial credit. These figures depict a fast and inexpensive model for agent execution, but not a universal champion of software development.
Security, partners, and first concrete uses
Meta says it evaluated Ember in its Superintelligence Scaling Framework for chemical and biological risks, cybersecurity, and loss of control. The model reportedly remained below thresholds deemed high risk in the tested categories, with safety training and tests similar to those of its other models. Behavioral evaluations also focused on respecting user intent, handling sensitive information, resistance to hidden instructions, and requesting human intervention before high-impact actions.
Early feedback cited comes notably from Cloudflare and You.com. Dane Knecht, CTO of Cloudflare, indicates Ember was up to 30% faster in tail latency and handled investigations covering multiple repositories well. At You.com, Chak Pothina reports a 95th percentile latency of 2.2 seconds and 36 to 48% fewer output tokens. For developers, the interest is obvious: delegate repetitive tasks to a less costly model and reserve the more expensive one for planning. It’s less glamorous than a chatty robot, but often more cost-effective.
Key points
- November 9, 2026: launch of Muse Ember by Meta.
- $0.10 per million input tokens.
- 36 points on the Artificial Analysis Intelligence Index.
- Cloudflare reports up to 30% gain in tail latency.
- You.com announces 2.2 seconds latency at the 95th percentile.
- Meta places Ember below Wonder and alongside Spark.
By the numbers
- 1 million tokens — context window announced by Meta.
- 512,000 tokens — maximum supported output volume.
- $0.14 per task — estimated cost of Ember on Artificial Analysis.
- $2.39 per task — estimated cost of Muse Spark 1.3.
- 2.2 seconds — 95th percentile latency reported by You.com.