A compact model designed for agents
Meta has announced Muse Ember, described by the company as its smallest and most efficient artificial intelligence model to date. According to a report by IT之家, the model is designed to perform complex reasoning tasks and the repetitive work that software agents carry out every day. The announcement therefore shifts attention from the benchmark score alone to two concrete aspects for application developers: the cost per individual operation and the ability to handle long sessions without switching models.
The key point is cost. Meta claims that, in the Artificial Analysis ranking, Muse Ember is on the same level as Claude Haiku 5.5 but at a cost per individual task of about one third; the company also maintains that the model outperforms GPT-6 Luna in several tests. These are the manufacturer's claims, reported by the source, and not independent verifications: the figures should be attributed to Meta and its evaluations.
The comparison figures and technical capabilities
The article lists the model's results against Claude Haiku 5.5 and GPT-6 Luna in various areas. Muse Ember scores 37.8 on AutomationBench-AA, a test of end-to-end business workflows, versus 13.9 and 20.2; 61.6 on Agents' Last Exam (57.0 and 57.1); and 46.8 on the Personal Superintelligence Index (36.1 and 37.0). In coding, it scores 40.9 on Terminal-Bench 4.0 (39.2 and 16.4) and 46 on the Coding Agent-AA index (37 and 31), while on FrontierCode 2.0-AA, which assesses the ability to integrate produced code, it stops at 27.2 versus 34.9 for Claude Haiku 5.5. On multimodal skills it scores 22.8 on GDP.pdf-AA, 69.5 on OSWorld 2.1 and 80.6 on PerceptionBench. In the overall Artificial Analysis Index, the two models are tied at 36, ahead of GPT-6 Luna at 28. The comparison therefore shows an uneven picture rather than clear superiority: Meta's model prevails in some operational tests, but not in all of them.
In terms of capabilities, Muse Ember has a one-million-token context window and can produce up to 512,000 output tokens, amounts that according to the source make it possible to process an entire code archive or a complete set of documents in a single call. The model is also described as natively able to understand text, images, audio and video, so as to read screenshots, charts and scanned documents, process voice and video recordings and operate the computer from a screen. These characteristics explain the orientation toward agentic use: a single model can follow multiple work steps without resorting to separate components.
Safety, availability and points still to be clarified
Meta states that, before release, Muse Ember was subjected to its SI evaluation framework and that all tests indicated a risk level below the high threshold set by the framework itself. This too is an internal company assessment, reported by the source, and not an external audit.
The model is available immediately through the Meta Model API, Google Cloud and Oracle Cloud. The listed price is 0.10 dollars per million input tokens and 0.50 dollars per million output tokens. It remains to be independently verified whether it is truly equivalent to Claude Haiku 5.5 and the cited cost ratio, as well as whether performance holds up in real applications compared with the tests presented.
