📰 Key Takeaways

Meta Superintelligence Labs released a new model, Muse Ember, on Monday, positioning it as the lab’s lightest and most cost-effective model to date. It’s built for executing complex reasoning and the work agents repeat all day: tool calling, coding subtasks, request routing, information extraction, and long-document parsing. Meta puts the spotlight on cost per task and cites Artificial Analysis’s composite benchmark score to claim Ember is cheaper and performs better than rivals like Anthropic’s Haiku and OpenAI’s Luna. It’s priced at 10 cents per million input tokens, well below comparable models from those frontier labs, and competitive with open-source models like DeepSeek V4.1 Flash and GLM 5.3 Flash.

In practice, Ember is meant to work alongside a larger model inside an agent framework, such as Meta’s Muse Wonder. Take a customer service agent: the heavier model writes the rules for handling a particular issue, then the lightweight model applies those rules across tens of thousands of cases. In Meta’s words, “Muse Wonder handles the planning, Muse Ember handles the volume.” Ember is also trained across text, images, and audio, so an agent can read screenshots, charts, and scanned documents directly, or work with voice memos, meeting recordings, and video, with no separate specialized model needed.

Ember is the fourth model under Meta’s Muse brand. The first three are the mid-tier Spark in April, the open-source Glimmer in August, and the flagship Wonder in October. The flagship consumer product, Muse Agent, launched in September with better-than-expected results, and earlier this month it picked up new features like Spaces and video chat.

💬 JudyAI Lab Take

Meta’s Muse Ember puts the selling point on cost per task. It’s priced at 10 cents per million input tokens, and Meta cites Artificial Analysis’s composite benchmark score to claim it’s cheaper and performs better than Anthropic’s Haiku and OpenAI’s Luna. In the lightweight model race, the focus is shifting from “who’s the strongest” to “who’s the best value.”

The design thinking here is that agent frameworks are moving toward division of labor. In Meta’s words, “Muse Wonder handles the planning, Muse Ember handles the volume”: the heavier model writes the rules for a customer service issue, and the lightweight model applies them across tens of thousands of cases. Work like tool calling, request routing, information extraction, and long-file parsing, the stuff agents repeat all day, doesn’t need a flagship model every single time. Another point worth noting is that Ember is trained on text, images, and audio, so an agent can read screenshots, scanned files, and meeting recordings directly without wiring in a specialized model, which cuts the number of moving parts in the system too. For AI builders, cost structure is now part of architecture design.

My suggestion: take inventory of your own agent workflows and mark which steps are just applying rules over and over. Try handing those steps to a lighter model, then compare per-task cost and quality.

📅 Source Info

🔗 Further Reading