Notes on film, new tools and the decisions behind the work.

Tools / Analysis

Meta shows Muse Ember porting 16 lexers with a larger model in charge

Ember handles repeated coding and document tasks at $0.10 per million input tokens at its standard rate. Meta’s examples show why the cost and correctness of the whole job matter more than that headline price.

An open aluminium folio gathers layered pages with a blue ribbon beneath MUSE EMBER lettering. Conceptual editorial illustration of coordinated work.
DU / SIGNALS · S/108Conceptual editorial cover · Google Flow

Sixteen workers take separate pieces of a Python codebase and translate them into Rust. A larger model prepares their assignments and checks the work; a smaller one carries out the repeated coding. That is the division of labour in one of Meta’s demonstrations for Muse Ember, a model the company introduced on 9 November for tasks that recur inside AI agents.

Ember is available through the Meta Model API and Muse Code as muse-ember-1.0. It accepts text, images, audio, video and PDFs and produces text. Its context window is 1,048,576 tokens, with a maximum output of 524,288 tokens, according to Meta’s documentation. A context window describes how much information can be supplied to a model; it does not mean the model will use every part equally well.

One planner, many bounded jobs

An agent may spend much of its time on routine operations: reading a file, looking up a field, editing one module or deciding where a request should go. If each step calls a large model, the cost of the workflow accumulates even when the difficult reasoning happens only occasionally.

Meta’s proposal is to assign those repeated steps to Ember, while Muse Wonder or Muse Spark handles planning and decisions that need more reasoning. The developer documentation is also explicit about the other side of that choice: for difficult single-shot coding, knowledge-heavy questions and visual coding, it recommends a larger model.

The Python-to-Rust example makes the split tangible. The task is to port 16 Pygments lexers: pieces of software that identify such elements as keywords, strings and comments in source code. Wonder plans and coordinates. Ember-based workers take the separate ports. Meta says the resulting lexers matched the Python reference across 383 test inputs and 62 held-out files from 18 public repositories, supplied only after the code was frozen. The company reports about $3.96 in model costs for this run, with results varying between runs.

That is a reported demonstration, not our own reproduction or a guarantee for an arbitrary migration. The reference implementations and checks matter: they give the workers a concrete target and make it possible to detect errors. A task with shared files, ambiguous requirements or no comparable reference could be harder to divide and verify.

Sources: Meta for Developers [2] · Meta for Developers [3]

What multimodal means here

The input list is broader than source code. Meta says Ember can work with images, video, audio and PDFs as well as text; its output is text. This is multimodal understanding within a workflow, not a claim that Ember generates films or finished images.

In another example, Meta describes sorting 10,000 photographs. Wonder selects categories from a sample, writes a pipeline and starts 16 Ember workers. The requests contain eight photographs at a time; Ember returns captions and album choices, which the pipeline uses to rename and file the images. Meta says the resulting albums agreed with the closest available dataset categories for about 86% of the photos. That figure is useful precisely because “all files processed” and “all files organised correctly” are different measurements.

The company’s developer article gives two different totals for this photo run, so we are not treating either amount as a verified final bill. The workflow itself still illustrates the trade-off: throughput comes from splitting the collection, while the labels need a quality check appropriate to the collection and its owner.

Sources: Meta for Developers [2] · Meta for Developers [3]

Price depends on the job, not just the headline rate

For requests up to 200,000 input tokens, Meta lists $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. Above 200,000 input tokens, the listed rates rise to $0.30, $0.012 and $1.50 respectively. The long-context threshold matters for an agent that repeatedly sends a growing conversation or a large document: a one-million-token window is available, but using more of it changes the price tier.

The higher tier applies to the whole request, not just the tokens above 200,000. Cached input and the history of a stored conversation count toward that threshold. A workflow that keeps appending context can therefore change its bill even if each worker’s new assignment remains short.

The same distinction applies to benchmark comparisons. On 10 November, Artificial Analysis listed Ember at maximum reasoning effort with 36 points on Intelligence Index v5.0.1 and an estimated cost of about $0.14 per index task. That cost reflects token use on the benchmark workload; it is not the price of every request or a prediction for a different workflow. Meta’s developer post gives about $0.13, while the live Artificial Analysis page showed $0.14 when checked.

Meta’s published table compares Ember, GPT-6 Luna and Claude Haiku 5.5 at maximum reasoning effort. Ember scores 37.8% on AutomationBench-AA against Luna’s 20.2% and Haiku’s 13.9%. On FrontierCode 2.0-AA, Ember’s 27.2% trails Haiku’s 34.9% and is above Luna’s 25.2%. These results describe the stated test setups; they do not establish a universal ranking across tasks or agent environments.

Sources: Meta for Developers [4] · Artificial Analysis [5] · Meta AI Research [1] · Meta for Developers [2]

IMAGE / PRODUCT

Intelligence Index score and estimated cost per task

Meta chart comparing Intelligence Index scores with estimated cost per task for Muse Ember, Claude Haiku 5.5 and GPT-6 Luna.
Credit / Meta for Developers / Artificial Analysis dataOpen original 

Meta’s published comparison uses Artificial Analysis results dated 9 November 2026 and list pricing for Ember. Each point represents a reasoning setting; the estimated cost belongs to the benchmark workload, not every use of the model.

Putting the split to work

Ember gives developers a way to spend the larger model’s effort on planning, then repeat narrower work at a lower token price. The savings depend on how well the task can be split: clear ownership, limited shared context and a result that can be checked keep the arrangement useful.

The code migration has a reference implementation and token-level checks. The photo job has a category-agreement measure. Those checks are part of the design, alongside the agents. To evaluate Ember on a real project, count planning, workers, retries and review together, and compare the finished result against a correctness target chosen before the run.

Sources: Meta for Developers [2] · Meta for Developers [3]

SOURCES & METHOD

The record behind the article

Reviewed 10 November 2026. Availability and pricing come from Meta’s documentation; the migration and photo-sorting results are Meta’s reported runs. Artificial Analysis provides the independent index result. Signals has not reproduced these workflows or run its own model benchmark.

  1. Introducing Our Smallest, Most Efficient Model: Muse Ember Meta AI Research · 9 November 2026
  2. Start building with Muse Ember, our smallest, most efficient model Meta for Developers · 9 November 2026
  3. Muse Ember — Meta Model API Meta for Developers
  4. Pricing and rate limits — Muse Ember Meta for Developers
  5. Muse Ember (max): intelligence, performance and price analysis Artificial Analysis