From the source
Meta introduced Muse Ember, an entry-level AI model it says costs significantly less than rival offerings and is optimized to run agent tasks inside ensembles anchored by a frontier model.
- Meta says Muse Ember matched
Anthropic's entry-level Claude Haiku 5.5 on the Artificial Analysis Index at about a third the price, and outperformed
OpenAI's comparable GPT-6 Luna.
- The same internal evaluation ran all three models in "max" mode across nine other benchmarks, where Muse Ember beat Haiku and Luna on seven.
- Its widest margin came on AutomationBench-AA, a set of 657 tasks spanning cloud applications such as Gmail and Salesforce, where it led by 17%, while most other results were only a few percent better.
- Muse Ember carries a 1 million token context window, accepts image, audio and video prompts, and can generate responses up to 512,000 tokens long.
- It is available through Meta's Model API at $0.1 per million input tokens and $0.5 per million output tokens, dropping to $0.004 per million input tokens when its built-in cache reuses frequent prompts.
- Meta says Muse Ember is the first AI it trained using Muse Wonder, its flagship large language model, though it did not specify which part of the workflow was automated.