Meta Superintelligence Labs has launched Muse Ember, a high-efficiency multimodal reasoning model engineered specifically for high-volume agent subtasks like coding steps, tool calls, and document parsing. Ember is the fourth model in Meta’s Muse family and features a 1-million-token context window alongside an output capacity of up to 512,000 tokens. It natively processes text, images, audio, and video inputs, allowing agents to execute tasks like operating desktop applications directly from screenshots.
Meta priced Muse Ember aggressively at $0.10 per million input tokens and $0.50 per million output tokens, matching list rates of competitor models such as Claude Haiku 5.5 and GPT-6 Luna. However, benchmark data from Artificial Analysis indicates that Ember performs equivalent tasks at approximately one-third the total cost of Haiku 5.5 due to token usage efficiency and the absence of high-tier pricing for long contexts.
The model was trained using Meta’s Muse Wonder Full model and is integrated into the Meta Model API and Muse Code developer environment. Early platform testers include enterprise partners Cloudflare, Oracle, Box, Cline, and You.com.
Why it matters
Sub-agent orchestration costs drop significantly with low-cost, long-context reasoning models focused on subtasks.
Native multimodal inputs (audio, video, vision) eliminate the latency and cost of separate pre-processing models.
Lack of long-context price surcharges creates aggressive price competition for enterprise document and codebase workflows.
Source: marktechpost.com



