Meta: Muse Ember

Meta introduced Muse Ember as a cost-efficient model for frequently repeated agentic tasks, including document processing, tool calls, and small programming jobs. Meta envisions stronger models such as Muse Wonder planning tasks while Ember executes the smaller steps, potentially reducing agent operating costs. Muse Ember accepts text, images, audio, video, and PDFs while generating text, supports about 1 million tokens of context, and can return up to 512,000 tokens. Meta says it was trained using Muse Wonder Full, though its architecture and weights remain undisclosed; it scored 36 on the Artificial Analysis Intelligence Index.

Meta released Muse Ember on November 9, 2026, positioning it as a lightweight, cost-efficient model for high-volume agent execution rather than top-level planning. The model ID is muse-ember-1.0, with a 1 million-token context window and a 512,000-token output ceiling. Meta prices Muse Ember at $0.10 per million input tokens and $0.50 per million output tokens, with access through Muse Code and the Meta Model API. Its intended workloads include tool calls, coding subtasks, routing, extraction, and parsing long documents or collections of media and records.

Meta Superintelligence Labs released Muse Ember on November 9, 2026, describing it as the lab's lightest and most cost-efficient model, purpose-built for complex reasoning execution and the repetitive work agents do all day, including tool calls, coding subtasks, routing, extraction, and long document parsing. Ember is priced at 10 cents per million input tokens, which Meta argued is cheaper than frontier rivals such as Anthropic's Haiku and OpenAI's Luna while remaining competitive with open-source models like DeepSeek V4.1 Flash and GLM 5.3 Flash, using Artificial Analysis composite scores. It also handles text, image, and audio inputs, and Meta positions it to pair with the larger Muse Wonder, where Wonder plans and Ember handles volume.

Hacker News discussion links to Meta's own announcement, titled "Muse Ember: Our smallest, most efficient model," confirming Meta as the creator of the model and positioning it as an efficiency-focused release. The thread attracted heavy developer participation, with 327 points and 139 comments, indicating broad interest in the launch. Simon Willison's comment documents six reasoning-effort levels for Muse Ember: minimal, low, medium, high, xhigh, and max. In his SVG pelican test, higher reasoning-token counts did not visibly improve output quality, breaking a correlation he had seen previously, and one level failed to render a complete SVG at all.

Vals AI’s independent Muse Ember evaluation places the Meta model 13th of 36 on its Vals Index, with a score of 60.48% and a $0.153 cost per test. The benchmark page lists a 1 million-token context window, 524,000 maximum output tokens, and input and output prices of $0.10 and $0.50 per million tokens. Muse Ember supports text, image, and file input but not video, and produces text output. Its listed benchmark results include 75.69% on the Vals Legal Index, 63.83% on Finance Agent v2.1, and 83.70% on Vibe Code Bench v1.1.

Meta’s Muse Ember launched November 9, 2026 as a low-cost model for repeated agent steps such as tool calls, routing, extraction, and quick code fixes. Its standard price is $0.10 per million input tokens and $0.50 per million output tokens, but prompts above 200,000 tokens cost $0.30 and $1.50 per million. The review reports that Muse Ember scored 36 on the Artificial Analysis Intelligence Index while costing $0.14 per task, versus $0.41 for Claude Haiku 5.5 with the same score. It won seven of Meta’s ten comparison rows, tied two, and lost one, although Claude Haiku led FrontierCode 2.0 by 34.9% to 27.2%.