Meta Launches Muse Ember, Its Smallest and Cheapest Model for AI Agent Work

Meta released Muse Ember on November 9, 2026, a small reasoning model built for the repetitive work AI agents do all day. The model handles tool calls, coding subtasks, routing, extraction, and long document parsing, and Meta calls it the smallest and most efficient model in its Muse family.

Ember is aimed squarely at volume. AI agents spend most of their tokens on many small, structured steps, and Meta designed Ember for exactly that kind of work. In its announcement, Meta says the model can carry a business workflow end to end, pulling records, calling tools in sequence, and updating systems across apps. The pitch is not raw capability but cost per task: Ember matches Claude Haiku 5.5 on a key benchmark at roughly a third of the cost.

A small model with a large context window

Ember offers a 1 million token context window and up to 512,000 output tokens, so it can take in a full codebase or document set in a single call. It natively understands text, images, audio, and video, which means agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without switching to a separate model.

The model is available now on the Meta Model API and in Muse Code, Meta’s coding agent, under the model ID muse-ember-1.0. It supports the Responses API. Meta also lists Ember on Google Cloud and Oracle Cloud. The weights are not open: unlike the open-weight Muse Glimmer, Ember is available only through hosted APIs.

Pricing starts at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Meta says the cost works out to about $0.14 per Artificial Analysis task, a figure it contrasts with $2.39 per task for its larger Muse Spark 1.3 model.

Benchmark claims put Ember next to Haiku 5.5

Meta leans heavily on benchmark results from Artificial Analysis, published the same day as the launch. On the Artificial Analysis Intelligence Index, Ember scores 36 at its maximum reasoning setting, the same score as Claude Haiku 5.5 and ahead of GPT-6 Luna at 28. Meta says no model scores higher for less per task, placing Ember on what it calls the efficiency frontier of the full leaderboard.

The detailed results are mixed in a way that matters for buyers. Ember leads Haiku 5.5 on AutomationBench business workflows (37.8 versus 13.9), on Terminal-Bench agentic coding (40.9 versus 39.2), and on the Coding Agent Index (46 versus 37). It trails Haiku on FrontierCode 2.0, which tests real engineering tasks judged on mergeability, scoring 27.2 against 34.9. These are vendor-reported figures from Meta’s announcement and from Artificial Analysis, not independent tests, so they should be read as launch claims rather than settled results.

Meta also says Ember is the first model it trained using Muse Wonder Full, the largest version of its frontier model. The suggested pattern is a division of labor: a large model plans, and Ember handles the volume. As Meta puts it in the announcement, “Muse Wonder plans, Muse Ember handles the volume.” TechCrunch gives the example of a customer support agent where a heavier model sets the rules for handling an issue and a lighter model then applies those rules across tens of thousands of cases.

Early partners and safety checks

In early testing, Meta’s partners put Ember to work on contract review and code writing. Box’s VP of AI Products, Yashodha Bhavnani, said the model can score a contract against a company’s negotiation playbook clause by clause and land on the right overall risk rating, or screen candidates against a detailed brief and flag the borderline ones who still qualify as alternates.

Before release, Meta evaluated Ember under its Superintelligence Scaling Framework, covering chemical and biological, cybersecurity, and loss-of-control risks. The company says the model stayed below the framework’s high-risk thresholds in every category assessed. Meta also tested how Ember behaves when acting on users’ behalf, including whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages or tool outputs, and seeks human input before high-impact actions.

Why it matters

Ember is the fourth model under Meta’s Muse brand, following the mid-weight Spark released in April, the open-weight Glimmer released in August, and the frontier Wonder model released in October. Together they form a tiered lineup: Wonder for the hardest problems, Spark for coding and agentic workflows, and Ember for repeatable work at scale.

The release reflects where the economics of AI agents are heading. As agents move from demos to production, the cost of the thousands of small steps they take each day becomes the deciding factor, and a model that is good enough at a fraction of the price can win more real work than a stronger but far more expensive one. Ember’s real test will be whether its benchmark parity with Haiku 5.5 holds up in production workloads, and whether developers trust a closed, API-only model from Meta for the high-volume layer of their agents. Meta says more efficient models are already in training.

Sources

• Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember, November 9, 2026
• TechCrunch: Meta releases Muse Ember, a new lightweight model for agent execution, November 9, 2026