Muse Ember

Meta Superintelligence Labs released Muse Ember on November 9, 2026, describing it as the lab's lightest and most cost-efficient model yet, aimed at tool calls, coding subtasks, routing, extraction, and long document parsing. Meta's framing pairs the lightweight Ember with the larger Muse Wonder in an agentic harness, summarized as "Muse Wonder plans, Muse Ember handles the volume." Base pricing is 10 cents per million input tokens, well below frontier competitors and roughly cost-competitive with open source models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta used Artificial Analysis composite benchmark scores to argue Ember is both cheaper and smarter than rivals like Haiku and Luna, and the model was trained to handle text, images, and audio inputs, including screenshots, charts, scanned documents, and voice memos.

OpenRouter lists Meta's Muse Ember as a model released on November 9, 2026, designed for complex reasoning and repetitive agent work such as tool calls, coding-agent subtasks, routing, extraction, and long-document parsing. The listing says it accepts text, images, video, audio, and PDFs, returns text, and provides a 1 million-token context window. The same listing reports $0.10 per million input tokens and $0.50 per million output tokens, alongside 163 tokens per second and 2.11-second median latency. It also lists a 5.83% tool-call error rate and a 1.56% structured-output error rate, though these are third-party hosted measurements rather than an official Meta model card.

Google Cloud's Gemini Enterprise Agent Platform documentation lists Muse Ember 1.0 from Meta as an efficient reasoning model built for high-volume agent tasks, including tool calls, coding subtasks, routing, extraction, and long-document processing. The exact version meta/muse-ember-1.0 is listed with a launch stage of Preview and a release date of November 9, 2026. Requests to the global /locations/global/publishers/meta/v1/responses endpoint pass the bare model ID muse-ember-1.0. The doc lists text and image inputs with text outputs, a maximum input of 1,048,576 tokens and maximum output of 524,288 tokens, plus preview-stage function calling, structured output, and reasoning support, with batch and provisioned throughput not supported.

Vals AI's independent benchmark profile lists Muse Ember as a Meta model released on November 9, 2026 with a 1M-token context window, up to 524k maximum output tokens, and private weights. Pricing is listed at $0.10 input and $0.50 output per million tokens, with text, image, and file inputs supported and video input not supported. On the Vals Index the model scores 60.48% ± 1.07, ranking 13th of 36 models, with a cost per test of $0.153 and latency of 15 minutes 2 seconds. Its strongest showing is second of 37 on HLAB-Vals-1.1 at 25.00%, while the Vals Legal Index reaches 75.69% (4th of 37), Finance Agent v2.1 reaches 63.83% (15th of 39), and Terminal-Bench 4.0 lands at 39.39% (16th of 63).

Meta launched Muse Ember on November 10, 2026 as its smallest and most efficient model, built for the repetitive work agents do all day: tool calls, coding subtasks, routing, extraction, and long document parsing. The post quotes Meta's announcement directly and names the variant Muse Ember 1.0, with a 1M-token input limit shared with competitors GPT-6 Luna and Claude Haiku 5.5. Ember is priced at $0.10 per million input tokens and $0.50 per million output up to 200k prompts, after which rates step up to $0.30 and $1.50, placing it between Luna and Haiku on long contexts. It also exposes six reasoning-effort levels, from minimal through max, and Meta says it sits on the Artificial Analysis efficiency frontier where no model scores higher for less cost per task.