Meta’s Muse Ember turns cheap agent work into the next AI battleground

Quick take

Meta introduced Muse Ember as its smallest and most efficient model, pricing it at 10 cents per million input tokens and 50 cents per million output tokens.
Meta’s developer documentation describes Ember as multimodal for text, images, audio and video, and recommends pairing it with Muse Wonder and Muse Spark.
Meta says Ember is cheaper on a composite cost-per-task measure than Anthropic’s Haiku and OpenAI’s Luna, while also competing with DeepSeek V4.1 Flash and GLM 5.3 Flash.
Plotly found Ember tied GPT-6 Luna on 38 of 43 data analysis questions, but calculated test costs of $1.04 versus Luna’s $0.13 and reported four missed hidden-pattern questions.

Why it matters

For companies building agents, Ember shifts the cost question from a model’s advertised token price to the expense and reliability of completing an entire workflow, especially when tasks are repeated at scale. Developers will need to see whether its multimodal, low-cost role reduces integration burdens without multiplying errors; the article says repetition, accuracy and predictable costs are the next test.

Most AI systems are built to impress people. Meta’s Muse Ember is built to keep busy, handling the repetitive digital work that agents must perform thousands or millions of times.

That work is rarely glamorous. It involves calling software tools, moving information between systems, extracting fields from documents, routing requests to the right model and completing small coding tasks. Yet as companies build agents that operate continuously, these routine actions may become one of the largest sources of AI spending.

Plotly benchmark cost: Ember $1.04 vs Luna $0.13; tied 38/43, Ember missed 4/5 hidden patterns. · Chart: SPAISEE · Data: plotly.com — see the chart in the original article.

Meta is betting that the best model for this work is not the most powerful one. It is the one that is inexpensive, fast enough and dependable when used at high volume.

In its announcement, Meta introduced Muse Ember as its smallest and most efficient model, targeting what it describes as high-volume agent work. The company lists tool calls, coding subtasks, routing, extraction and long-document processing among its intended uses. Ember is priced at 10 cents per million input tokens and 50 cents per million output tokens, according to Meta.

Meta Muse Ember listed price per million tokens — see the chart in the original article.

That pricing is designed for a different role from a flagship model. A large system such as Muse Wonder might plan a complicated task, make a judgment or communicate with a customer. Ember is intended to carry out the many smaller steps beneath it. The distinction resembles the difference between a senior manager and an operations team. One decides what needs to happen. The other makes sure the process keeps moving.

The economics of the agent stack

For businesses, this division could matter more than another small improvement on a public benchmark. An agent that reads a document, checks a database, calls an application programming interface and updates a record may use a model repeatedly during a single workflow. At that scale, even fractions of a cent can shape whether automation is profitable.

Meta says Ember is cheaper on a composite cost-per-task measure than Anthropic’s Haiku and OpenAI’s Luna. It also positions the model against other low-cost systems, including DeepSeek V4.1 Flash and GLM 5.3 Flash. Those comparisons are part of Meta’s own presentation, rather than an independent industry assessment, but they show where the company sees the contest taking place.

The competition is therefore moving below the level of headline model launches. The important question is not only which system can produce the most impressive answer. It is which system can complete a mundane operation reliably enough that a company is willing to entrust it with a constant stream of work.

That creates a difficult tradeoff. A cheaper model can reduce the cost of every action, but mistakes can multiply just as quickly as successful calls. An incorrect extraction from one invoice may be manageable. The same error repeated across thousands of documents can become an accounting problem, a customer service burden or a compliance risk.

One model for more than text

Ember’s other distinguishing feature is its range of inputs. Meta’s developer documentation describes Muse Ember as a multimodal model for text, images, audio and video, and recommends using it alongside larger models such as Muse Wonder and Muse Spark in agent workflows.

That means an agent could potentially work from a screenshot, a scanned form, a chart, a recording or a video instead of sending each format to a separate specialist model. For developers, the attraction is simplicity. Fewer model handoffs can mean fewer integrations, fewer opportunities for information to be lost and a more straightforward system to maintain.

It also brings the model closer to the messy conditions of ordinary work. Business information does not arrive neatly as text. It appears in photographs, attachments, meeting recordings, spreadsheets and old documents that were never designed for machine processing. A lightweight system that can interpret several formats could be useful precisely because it fits into those imperfect workflows.

Meta’s own recommendation to pair Ember with larger models reveals the intended architecture. Ember is not necessarily meant to replace the strongest systems. It is meant to sit underneath them, performing routine tasks while a more capable model handles difficult reasoning or final decisions.

Cheap does not always mean efficient

The early independent reaction included a warning about how headline pricing can differ from the cost of completing a real job.

Plotly’s independent benchmark found that Ember tied GPT-6 Luna on 38 of 43 data analysis questions, but used substantially more tokens in the process. Plotly calculated the cost of the test at $1.04 for Ember, compared with $0.13 for Luna, despite the same score. The company also said Ember missed four of five questions involving hidden patterns.

That result does not invalidate Meta’s input-token price. It points to a broader issue: the cheapest token is not always the cheapest completed task. If a model needs lengthy internal reasoning, repeated attempts or more output to reach an acceptable answer, the bill can rise even when the listed rate is low.

Spring Prompt reached a similarly critical conclusion in its own testing. Spring Prompt reported that Ember was strong at analysis but less attractive on a per-job basis because of its reasoning-token use. Its review said Luna performed better for sales copy and slide decks, while Ember was slower on short responses and did not offer open weights.

These tests represent criticism, not a complete verdict. They examine particular tasks, and performance can change depending on how an agent is designed, how much context it receives and whether the job requires deep reasoning. Still, they identify the central risk in Meta’s strategy. A model intended for volume must be measured by the cost and reliability of the entire workflow, not by its advertised token rate alone.

The real test is repetition

Only critical outside assessments were found for this launch, so there is not yet a documented opposing view from independent reviewers in the supplied material. Meta’s case rests on a straightforward proposition: agent systems will need a large number of inexpensive workers, and Ember can provide them across several data types.

The next stage will be less about whether Ember can produce a convincing demonstration. It will be about whether it can perform the same narrow task repeatedly, with few errors and predictable costs. That is a more demanding standard than answering a single question.

If Ember meets it, lightweight models could become the hidden infrastructure of AI products, handling the invisible work that makes larger systems useful. If it does not, companies may discover that saving money on each call is less important than avoiding the cost of correcting thousands of small mistakes.