Meta Launches Muse Ember: Lightweight Agent Model at 10¢ per Million Tokens

Executive Overview

Meta Superintelligence Lab announced on Monday the launch of Muse Ember, its lightest and lowest-cost model to date, priced at 10 cents per million input tokens. The model is engineered for high-volume, repetitive tasks that agents execute daily, and it operates in tandem with the heavier Muse Wonder model within an agentic framework. According to Meta, Ember is smarter and cheaper per task than Anthropic Haiku and OpenAI Luna on the Artificial Analysis composite benchmark, while remaining cost-competitive with DeepSeek V4.1 Flash and GLM 5.3 Flash. As the fourth model in the Muse family, Ember underscores Meta's push toward cost-per-task economics in agentic AI.

Deep-Dive Features & Architecture

Muse Ember is built around a cost-per-task philosophy rather than cost-per-token alone. At 10 cents per million input tokens, it targets the repetitive tasks agents perform throughout the day: tool calling, sub-coding tasks, routing, data extraction, and long-document analysis. Meta asserts that Ember is smarter and cheaper per task than Anthropic Haiku and OpenAI Luna on the Artificial Analysis composite benchmark, with competitive costs against DeepSeek V4.1 Flash and GLM 5.3 Flash.

Ember operates within a plan-and-execute architecture alongside Muse Wonder: the heavier model sets the rules and processing plans, and Ember executes those rules across tens of thousands of instances. Meta summarizes this as “Muse Wonder plans, Muse Ember handles the volume.” Ember is also trained on multiple media types—text, image, and audio—enabling agents to read screenshots, charts, and scanned documents, and to work from voice notes, meeting recordings, and video without a separate model.

Ember arrives as the fourth model in the Muse family, following the mid-weight Spark (April), its open-source twin Glimmer (August), and the advanced Wonder model (October). The consumer Muse Agent product launched in September with unexpected success, alongside Spaces and video chat features rolled out in early November.

Benchmark & Competitive Performance

According to the Artificial Analysis composite benchmark, Muse Ember outperforms Anthropic Haiku and OpenAI Luna on cost per task, positioning it in direct competition with low-cost open-source models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta relies on this composite benchmark rather than raw speed metrics because the market is shifting toward measuring actual value per task executed, not just token cost. This approach makes Ember an economical choice for enterprises running agents at scale, where cost-per-task differences compound under high-volume processing.

Industry Impact & Enterprise Adoption

For developers in the Arab world, the 10 cents per million input tokens price point marks a turning point in the economics of building agent applications, especially given that Arabic text typically consumes 30% to 60% more tokens than English. This means the cost of processing long Arabic documents or meeting recordings will remain relatively low compared to advanced models. Ember's support for audio and visual inputs also opens practical use cases in the Arab market, such as analyzing Arabic scanned documents, processing voice notes, and automating customer service across multiple channels.

Conclusion

Muse Ember represents Meta's strategic bet on cost-efficient, high-volume agent execution. By pairing a lightweight, multimodal model with the heavier Muse Wonder in a plan-and-execute framework, Meta offers enterprises a scalable path to deploy agents without prohibitive token costs. As the AI industry shifts toward measuring value per task, Ember's pricing and benchmark positioning could set a new standard for economical agentic AI.

Media Source: TechCrunch AI | Fact Verification & Analysis: AI Tools Oasis