Meta introduces its smallest model, Muse Ember, for agents’ day-to-day tasks
On November 9, 2026, Meta Superintelligence Labs released its smallest and most efficient model to date, Muse Ember, designed for complex reasoning execution and agents’ routine repetitive work, including executing workflows across apps, calling tools sequentially, running tests, and making targeted code fixes.
The model natively supports text, image, audio, and video inputs, with a context window of 1 million tokens and a maximum output of 512,000 tokens. Pricing is $0.10 per million input tokens, $0.50 per million output tokens ($0.004 for cached input), and it is now available through Muse Code and the Meta Model API (model ID muse-ember-1.0).
Meta said that Muse Ember matched Claude Haiku 5.5’s top score of 36 points on the Artificial Analysis Intelligence Index at a cost of about $0.14 per task, roughly one-third of the latter’s cost, and outperformed GPT-6 Luna across all reasoning settings. In benchmarks such as Terminal-Bench 4.0 (40.9% versus 39.2%) and OSWorld 2.1 (partial credit 69.5% versus 68.9%), the company’s published results showed it scoring higher than the two models above; however, on FrontierCode 2.0-AA it was lower than Haiku 5.5 (27.2% versus 34.9%). The model is the first model trained using Muse Wonder Full.
Original sources:
Title: Introducing Our Smallest, Most Efficient Model: Muse Ember Published by: Meta AI Research Publication date: 2026-11-09 URL: https://research.meta.ai/blog/muse-ember Source type: official announcement Verification status: The full text of the official blog post, benchmark tables, pricing and availability notes were read in full for this report (2026-11-09) Supporting content: supports Muse Ember’s release date, positioning (agents’ daily tool calling, coding subtasks, cross-app workflows), multimodal input, 1M context and 512K output, pricing ($0.10/$0.50), launch channels (Muse Code and Meta Model API), the company’s claimed comparison results with Claude Haiku 5.5 and GPT-6 Luna on the Artificial Analysis Index, Terminal-Bench 4.0, OSWorld 2.1 and other benchmarks, and that it is the first model trained with Muse Wonder Full.
ABAB AI Insight
Meta Superintelligence Labs Chief AI Officer Alexandr Wang and Mark Zuckerberg both announced the release on X. Since the launch of the first Muse model, Muse Spark, in April 2026, the Muse series has been layered step by step: Spark is positioned for mid-tier coding and agent workflows, the open-weight Glimmer released in August is aimed at local running, the frontier model Muse Wonder was released in October, and Ember fills the high-capacity, low-cost execution layer. The company clearly said that Ember is the first model distilled or trained using Muse Wonder Full, intending to use the larger model for planning and the smaller model for repetitive volume work.
On the capital and product path, Ember does not involve a new round of external financing, but rather the completion of Meta’s internal product matrix. Meta aligned its pricing with Haiku 5.5 and Luna (both $0.10 per million input tokens and $0.50 per million output tokens), but by using fewer output tokens and lower cached-input costs, it reduced the Artificial Analysis cost per task to about $0.14, far below Haiku’s $0.41. The flow of funds is reflected in developers using the API or Muse Code to call it for sub-agent tasks, tool loops, and document processing, thereby reducing the marginal cost of the overall agent system.
Compared with Anthropic’s Claude Haiku series, both are positioned as efficient execution layers, and Haiku has long served as Anthropic’s low-cost option for coding and tool calling. The key similarity is that the list prices are close and both emphasize repetitive agent tasks; the difference is that Meta proactively disclosed cost-per-task data and claimed to be at the frontier of Artificial Analysis efficiency, while clearly designing Ember to be used together with Wonder/Spark rather than as a standalone main model. Historical cases show that similar layering, such as OpenAI’s o series and smaller models, often pushes agents from experimentation into high-frequency production, but it depends on actual output-token efficiency and tool reliability.
This release changes the cost structure and model selection for agent developers: high-volume subtasks can be offloaded from larger models to Ember, potentially lowering per-session token spending, especially in cross-app workflows and long-document scenarios. The key metrics to watch are output-token consumption and task completion rates in real production; if they continue to stay below competitors, this will strengthen Meta’s price competitiveness in the agent execution layer, provided that the advantages seen in benchmarks can be reproduced in real harnesses and that the higher pricing for long context (more than 200,000 tokens) does not offset the overall savings.
ABAB News · Cognitive Laws
Having the same list price does not mean the same per-task cost.
Small models handle volume, large models handle planning; this is a common path for agent layering.
Benchmark advantages must be verified in actual tool loops before they can translate into cost advantages.