Meta's Muse Ember matches Claude Haiku 5.5 at a third the price

Meta's pitch for Muse Ember rests on a specific architectural role rather than raw capability. Applications increasingly use frontier LLMs such as Muse Wonder as part of AI ensembles: the frontier model splits a user request into smaller tasks and hands them to agents powered by a lower-cost model. Meta says Muse Ember is built for that second role, which explains why the company emphasizes price parity with Claude Haiku 5.5 over outright benchmark leadership.

The evaluation details Meta shared point the same way. On most benchmarks, Muse Ember's results were only a few percentage points better than Haiku and Luna; the outlier was AutomationBench-AA, a collection of 657 tasks spanning popular cloud applications such as Gmail and Salesforce, where it recorded a 17% lead. Whether that gap holds outside Meta's internal testing is not established by the announcement.

Muse Ember is also notable as the first AI Meta trained using Muse Wonder, its flagship model, though Meta did not specify which part of the workflow was automated. The company's stated capabilities — a 1 million token context window, image, audio and video inputs, and responses up to 512,000 tokens — target lengthy business documents rather than consumer chat.