Meta Releases Lightweight Muse Ember Model at One-Third Task Cost
Meta has launched Muse Ember, its smallest and most efficient model to date, purpose-built for complex reasoning and high-frequency agentic workloads. It matches Claude Haiku 5.5 on the Artificial Analysis Index (36) at one-third the per-task cost, while leading on agentic and multimodal benchmarks such as AutomationBench-AA and OSWorld 2.1. The model is available immediately via Meta Model API, Google Cloud and Oracle Cloud, priced at $0.10/$0.50 per million input/output tokens.
Full Report
IT之家 November 10 news: Meta has now launched the Muse Ember AI model. This is Meta's smallest and most efficient model to date, designed for complex reasoning task execution and the repetitive day-to-day work handled by agents.
Meta states that on the Artificial Analysis leaderboard, this model's performance rivals Claude Haiku 5.5, but its cost per task is only 1/3, and it surpasses GPT-6 Luna in multiple tests.
IT之家 attaches the Muse Ember model benchmark data as follows:
| | Benchmark | Muse Ember (max) | Claude Haiku 5.5 (max) | GPT-6 Luna (max) | | --- | --- | --- | --- | --- | | Agentic | AutomationBench-AA (end-to-end business workflows) | 37.8 | 13.9 | 20.2 | | Agents' Last Exam (end-to-end agentic tasks) | 61.6 | 57.0 | 57.1 | | | Personal Superintelligence Index (internal composite index for personal and knowledge work) | 46.8 | 36.1 | 37.0 | | | Coding | Terminal-Bench 4.0 (agentic terminal coding) | 40.9 | 39.2 | 16.4 | | Coding Agent Index-AA (coding in real agentic environments) | 46 | 37 | 31 | | | FrontierCode 2.0-AA (real engineering tasks evaluated by code merge feasibility) | 27.2 | 34.9 | 25.2 | | | Multimodal | GDP.pdf-AA (long-document professional work) | 22.8 | 20.8 | 22.8 | | OSWorld 2.1 (computer operation across desktop applications) | 69.5 | 68.9 | 50.5 | | | PerceptionBench (fine-grained visual perception) | 80.6 | 78.6 | 62.8 | | | Overall | Artificial Analysis Index | 36 | 36 | 28 |
At the same time, this model has a 1-million-token context window and supports up to 512K output tokens, allowing it to process an entire codebase or a full set of documents in a single call. It enables an agent to execute hundreds of coding subtasks in a single session, including reading files, running tests, and making targeted fixes.
Muse Ember natively supports understanding text, images, audio, and video. Without relying on other models, the agent can read screenshots, charts, and scanned documents, and can also process voice memos, meeting recordings, and videos. It can handle lengthy real-world documents, operate a computer directly based on screen captures, and recognize subtle visual details.
In terms of safety, this model was evaluated under Meta's SI extended framework before release, and all test results show that Muse Ember's risk level is below the high-risk threshold set by the framework.
In terms of pricing, this model is available starting today on the Meta Model API, Google Cloud, and Oracle Cloud. The input price is $0.10 per million tokens (IT Home note: approximately 0.67 yuan at current exchange rates), and the output price is $0.50 per million tokens (approximately 3.3 yuan at current exchange rates).
