Muse Ember: Meta's model designed for AI agents
Meta presents Muse Ember, a multimodal model tailored for repetitive AI agent tasks. Its one million token context window mainly aims to reduce execution costs.
Meta unveiled Muse Ember on November 9, 2026, an artificial intelligence model designed to automate repetitive AI agent tasks at scale. Intended for tool calls, data extraction, document analysis, and certain programming tasks, it focuses on cost reduction rather than competing for the most powerful model. The idea is simple: reserve the heavy brains for complex problems and let Ember handle the groundwork.
A model built for execution, not for show
Meta positions Muse Ember as an infrastructure component for companies wanting to industrialize agents without constantly running a high-end model. This specialization matters because agents often chain intermediate operations before reaching a useful action; at scale, every token consumed adds up on the bill.
The model is multimodal: it processes text, images, audio, and video. Meta also states it can analyze screenshots, scanned documents, and meeting recordings without systematically relying on another model. Ember is available via the Meta Model API, Muse Code, and the Responses API, under the identifier muse-ember-1.0.
A one-million-token context window
The main technical feature highlighted by Meta is its one-million-token context window, with a maximum generation of 512,000 tokens. In practice, this allows it to absorb large volumes of information in a single request, whether it’s a codebase or a substantial document set. Meta does not disclose the number of model parameters.
This capability targets use cases where an agent must read, cross-reference, then execute without losing track. In this context, Ember can serve as the execution engine, while the Muse Wonder model takes over planning and more demanding reasoning. Muse Spark remains focused on programming and agentic processes.
Competitive performance, if you look at the right chart
On the Artificial Analysis intelligence index, Muse Ember scores 35.8 at the highest reasoning level, compared to 36.2 for Anthropic’s Claude Haiku 5.5 and 27.8 for OpenAI’s GPT-6 Luna. On AutomationBench-AA, which evaluates 657 automated business processes, it achieves a 37.8% success rate, versus 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5.
Meta also claims 69.5% on OSWorld 2.1, using a partial credit evaluation method. These figures remain dependent on test conditions. Claude Haiku 5.5 also retains the advantage on FrontierCode 2.0, dedicated to software engineering, with a 34.9% success rate against 27.2% for Ember.
A price designed for high-volume use
The price announced by Meta is 0.10 dollars per million input tokens, 0.004 dollars for cached tokens, and 0.50 dollars per million generated tokens. These prices are identical to those listed for Claude Haiku 5.5 and GPT-6 Luna; Meta’s argument therefore does not rest on unit price but on the number of tokens needed to complete a task.
For requests with input exceeding 200,000 tokens, Meta applies higher rates. This pricing structure may interest companies wanting to automate processing chains with predictable costs. The real question remains, less glamorous than score tables: how many times will the agent have to restart the machine before doing exactly what is expected?
Key points
- Presentation on November 9, 2026, according to Meta.
- Muse Ember processes text, image, audio, and video.
- Context window: one million tokens.
- 37.8% on AutomationBench-AA, according to Meta.
- 0.10 dollars per million input tokens.
- Meta sees it as a complement to Muse Wonder.
By the numbers
- 1 million tokens — Muse Ember’s context window, announced by Meta on November 9, 2026.
- 512,000 tokens — maximum generation announced by Meta.
- 37.8% — success rate on AutomationBench-AA, according to Meta.
- 0.10 dollars — price per million input tokens via the Meta Model API.
- 657 processes — scope of AutomationBench-AA testing.