Meta releases Muse Ember: an agent model at $0.10 per million input tokens

If you have an agent running in production, the bulk of your bill usually doesn't pile up on the "smart" steps but on the boring ones: parsing a document, calling a tool, sending a request down the right route, pulling fields out of an email. There are thousands of such steps per hour, and running a flagship model on them is like calling in the chief engineer to change light bulbs. This is exactly the niche Meta is targeting with its new model, Muse Ember.

What Meta released

Meta Superintelligence Lab introduced Muse Ember on Monday. It is the company's lightest and most economical model. Its stated purpose is "the work agents repeat all day long": tool calls, coding subtasks, routing, data extraction and parsing of long documents.

The main selling point is price: 10 cents per million input tokens. That is noticeably cheaper than Anthropic's Haiku and OpenAI's Luna, and comparable to the open models DeepSeek V4.1 Flash and GLM 5.3 Flash. For its "price versus quality" comparison, Meta used the Artificial Analysis composite score and showed that Ember is not only cheaper but also smarter than Haiku and Luna.

Another trump card is multimodality. Ember was trained to work with text, images and audio. According to Meta, agents can read screenshots, charts and scanned documents, and also work with voice notes, meeting recordings and video without a separate model. If that holds up, the pipeline gets fewer links: separate speech recognition and OCR may no longer be needed.

What this looks like in practice

Meta itself proposes a pairing of two models. Muse Wonder, the flagship released in October, plans. Ember handles the volume. The formula from the announcement: "Muse Wonder plans, Muse Ember handles the volume."

Take customer support. The heavy model analyzes a type of request once, say a refund for a subscription, and formulates the rules: what to check, when to approve, when to hand off to a human. Then the light model applies those rules to tens of thousands of requests: it reads the email, looks at the error screenshot, triggers the right tool in the CRM, fills in the fields. The expensive model works rarely, the cheap one constantly, and it is the cheap one that determines the final bill.

The idea is not new: model cascades and routing have been used for a long time. But Meta offers a ready-made configuration of two models from the same family. Ember is the fourth in the Muse lineup, after the mid-size Spark (April), its open sibling Glimmer (August) and the flagship Wonder (October).

Caveats

All the figures come from Meta's announcement. The company chose the comparison metric itself, and the composite index averages different tasks, so your agent may behave differently. The material does not make clear how much output tokens cost (in agent loops they often weigh heavily), what the model's context window is, or whether the weights can be downloaded. Glimmer is described as open; nothing of the kind is said about Ember.

Another trap: price per token is not price per task. If the cheap model takes more steps or makes more mistakes and requires retries, the savings melt away. It is no accident that Meta talks specifically about cost-per-task.

What to do about it

If you have an agent with a large volume of repetitive steps, take real logs, run 500–1,000 tasks through Ember and through your current workhorse, and compare the cost of a successfully completed task, not of a token.

If you haven't yet split your agent into a planner and an executor, now is the time: even without Meta, it is usually the first optimization of the bill. If you use separate models for screenshots and audio, check whether a single one can handle them. And don't forget to look at output token pricing and limits once they appear: that is where a cheap model sometimes turns out not to be so cheap.