Meta announced Muse Ember, its smallest and most efficient model for agents. It reported benchmark performance exceeding that of comparable competitors despite pricing of $0.10 per 1M input tokens and $0.50 per 1M output tokens, in the same price range. The company aims for low-cost operations by combining it with the related large-scale model Muse Wonder.
Meta announced its new AI model Muse Ember, designed to handle routine repetitive tasks performed by agents at low cost. Muse Ember has a 1M-token context window and is said to be capable of directly handling text, images, audio and video. Pricing is $0.10 per 1M input tokens and $0.50 per 1M output tokens, and Meta said its performance exceeds that of competing models in the same price range.
According to the official announcement, Muse Ember became available as muse-ember-1.0 on Muse Code and the Meta Model API. It is said to support continuing workflows across applications, sequential tool calls, test execution and code modification. The design emphasises combining it with Muse Wonder, which serves as a larger command-centre model, to handle execution tasks at low cost.
The benchmark results showed Ember (maximum setting) at 37.8%, compared with 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5 on AutomationBench-AA. On Agents’ Last Exam, Ember scored 61.6%, compared with 57.1% for Luna and 57.0% for Haiku 5.5. The results were described as indicating that Ember has "higher intelligence than lightweight models in the same class."
Meta CEO Mark Zuckerberg positioned Ember as "the smallest model with the highest intelligence per dollar in its class." Muse Ember is intended to reduce costs in long-running agent operations together with Muse Wonder. On the same day as the announcement, Anthropic also announced Claude Fable 5.2 and Mythos 5.2, drawing attention to moves toward differentiated cost structures.