On November 9, 2026, Meta released Muse Ember (muse-ember-1.0) — the smallest and cheapest model in the Muse lineup, designed for high-volume agentic tasks: tool calls, coding subtasks, routing, and long-document processing. According to Artificial Analysis measurements, its intelligence is comparable to Claude Haiku 5.5 at approximately one-third the cost per task. The model is already available via Meta Model API, Muse Code, Google Cloud Model Garden, and Oracle Cloud.
What happened
Meta released a model with the identifier muse-ember-1.0, filling the lower price tier of the Muse lineup. It is multimodal: it works with text, images, audio, and video. The context window reaches 1 million tokens, with an output limit of up to 524,288 tokens. Pricing: $0.10 per 1M input tokens, $0.50 per 1M output tokens, and $0.004 per 1M cached input tokens, meaning the cache is 25 times cheaper than the base input. According to Artificial Analysis data from November 9, 2026, in max reasoning mode, the model scores 35.8 on the Artificial Analysis Intelligence Index at approximately $0.136 per task; Claude Haiku 5.5 shows a score of 36.2 at approximately $0.411 per task, while GPT-6 Luna scores 27.8 at $0.152. Integration is possible via Meta Model API and Muse Code, as well as in Google Cloud Model Garden with Preview status and a global endpoint, and in Oracle Cloud.
Context
In the Muse lineup, the top tier is occupied by frontier models Muse Wonder and Muse Spark 1.3, which remain for planning and complex decision-making, while Ember is assigned the role of a cheap executor for mass pipeline steps. The discounted rate for cached input, combined with long context, targets a specific workload: multi-step agentic cycles where the same context is actively reused rather than rebuilt. The competitive landscape describes market segmentation rather than displacement: if Artificial Analysis measurements are reproduced, Ember sets the standard in the segment of approximately $0.13–0.15 per task, Claude Haiku 5.5 remains the leader in absolute quality at a different price point, and GPT-6 Luna, at a price comparable to Ember, trails it by 8 points.
Why this matters for the industry
Mass steps in agentic pipelines — tool calls, data extraction and parsing, code fragments, routing — are transforming from an expensive line item into a cheap building block that can be delegated to an Ember-level executor, leaving frontier models for planning. The combination of 1M token context with long output changes the class of tasks solvable in a single call: it allows loading entire codebases and document arrays in full. If independent measurements are confirmed, the consolidation of two-layer architectures of the form "cheap executor plus expensive planner" is expected, and price pressure will shift to the Haiku-class model segment. As access to cheap executors becomes commoditized, product value shifts toward orchestration, proprietary data, eval quality, and integration surface.
Why this matters for users
The model can be tested right now: muse-ember-1.0 is accessible via Meta Model API and Muse Code, and is also open in Google Cloud Model Garden with Preview status and in Oracle Cloud. A reasonable first step is a pilot on non-critical tasks with a fallback to the current model and comparison with Claude Haiku 5.5 and GPT-6 Luna on your own workload with a fixed budget. For applied scenarios, this is a cheap way to automate routine tasks without third-party tools: processing screenshots, meeting recordings, and long documents, while cached input makes repeated runs on the same context cheaper.
What is currently unknown / limitations
Independent confirmation of quality is currently singular: this is the only Artificial Analysis measurement for one day on one index, without disclosed methodology in the provided materials; the second source on quality is Meta's own evaluation report, i.e., manufacturer data. Sources differ on the output limit: the post text mentions 512k tokens, while the summary states 524,288. The statement about outperforming competitors is true only in the price/intelligence metric: in absolute intelligence, Ember slightly trails Claude Haiku 5.5. Corporate access via Google Cloud remains in Preview, and before migrating loaded pipelines, the stated figures should be reproduced on your own workload.
Sources
- Introducing Our Smallest, Most Efficient Model: Muse Ember — Meta AI Research
- Muse Ember Evaluation Report — Meta AI Research
- Muse Ember 1.0 — Google Cloud Gemini Enterprise Agent Platform Documentation
Author
Look at AI, editorial team