Overview
Muse Ember is a small model from Meta Superintelligence Labs, released on November 9, 2026, aimed at the tier agents lean on for legwork: calling tools, writing short snippets, routing requests and pulling fields out of long documents. It is also the smallest and most intelligence-dense-per-dollar model Meta ships.
The boundaries are stated plainly. It carries a 1M-token context window, natively understands text, image, audio and video, and can run workflows across applications, call tools in sequence, execute tests and patch what fails. Those are precisely the actions an agent repeats hundreds of times a day, and the line item most likely to blow up a bill.
Artificial Analysis scores it 36 on the Intelligence Index, level with the top setting of Claude Haiku 5.5, at $0.14 per task, roughly a third of that model's cost. On Arena's Code Arena WebDev it reaches 1608, the best small model on the board, ahead of Claude Haiku 5.5 at 1595 and GPT-6 Luna at 1578. All three list the same $0.10 input and $0.50 output per million tokens, yet Ember's median cost lands at $0.03 per task against $0.06 for Luna and $0.11 for Haiku. Identical stickers, different token budgets.
The formal name is muse-ember-1.0. It is available today on the Meta Model API and in Muse Code, listed on OpenRouter, and Meta brought it to Google Cloud alongside Muse Wonder and Muse Spark 1.3.
Key Features
- 1M-token context: Handles a million tokens of documents in one pass, suited to long-document parsing, batch extraction and cross-file reasoning.
- Native multimodal input: Understands text, image, audio and video natively, so transcription and recognition models are not needed as a preprocessing step.
- Agent execution: Runs workflows across applications, calls tools in order, executes tests and patches failures, built for tool calling, coding subtasks and request routing.
- Intelligence per dollar: $0.10 input and $0.50 output per million tokens, with a $0.14 cost per task, the cheapest model reaching 36 on the AA Intelligence Index.
- Small-model benchmark standing: 1608 on Code Arena WebDev, the highest small model on the board, at a median $0.03 per task.
- Four ways to call it: Meta Model API, Muse Code, OpenRouter and Google Cloud, which pairs well with Muse Wonder for separating planning from execution.
Use Cases
- Front-end agents dispatching bounded subtasks to a cheap model while planning stays on a flagship.
- Batch parsing and field extraction across long documents.
- Request routing and intent classification where call volume is high every day.
- Tool calls and failure patching inside cross-application workflows.
Pros
- $0.14 per task matches Claude Haiku 5.5 at its top setting for about a third of the cost.
- A 1M context is rare in the small-model tier, so long documents avoid chunking.
- Native text, image, audio and video support removes preprocessing stages.
- 1608 on Code Arena WebDev puts it first among small models at a median $0.03 per task.
- Callable from Meta Model API, OpenRouter and Google Cloud, so integration cost is low.
Pricing
$0.10 per million input tokens and $0.50 per million output tokens, the same list price as Claude Haiku 5.5 and GPT-6 Luna. Median cost is $0.03 per task, while Artificial Analysis reports $0.14 per task. Available through the Meta Model API, Muse Code, OpenRouter and Google Cloud.
Summary
Muse Ember is Meta's small model for high-frequency agent work, with a 1M context, native multimodality and the ability to run tool calls and patch failures. At 36 on the AA Intelligence Index it matches Claude Haiku 5.5 at its top setting for about a third of the per-task cost, and 1608 on Code Arena WebDev makes it the leading small model at a median $0.03 per task. It fits the execution layer behind a planner, taking the repetitive work.
Version History
- Muse Ember launches as Meta's smallest efficient model (2026-11-09): Meta released muse-ember-1.0 on November 9, 2026 as its smallest complex-reasoning executor, built for tool calling, coding subtasks, routing, information extraction and long-document parsing, with a 1M-token context and native text, image, audio and video understanding. Pricing is $0.10 input and $0.50 output per million tokens. Artificial Analysis scores it 36 on the Intelligence Index, level with the top setting of Claude Haiku 5.5 at about a third of the per-task cost of $0.14; Code Arena WebDev puts it at 1608, first among small models, with a median $0.03 per task. It launched on the Meta Model API, Muse Code and OpenRouter, and arrived on Google Cloud alongside Muse Wonder and Muse Spark 1.3.