TabbitBlog

Muse Ember: Meta’s Small Model for Repeated Agent Tasks

See where Muse Ember fits in coding agents, how its API and Muse Code rollout work, and what the long-context pricing threshold changes.

In this article
  1. Key takeaways
  2. Muse Ember at a glance
  3. Choose Ember for repeatable subtasks, not every coding job
  4. A small evaluation you can run safely
  5. API access, Muse Code and the pricing threshold
  6. Where Tabbit fits, and where it does not
  7. Verdict: test Ember on repeatable, bounded work

Muse Ember is Meta’s hosted model for repeated, bounded work inside an agent: focused coding subtasks, tool calls, routing, extraction and document processing. It is a reasonable candidate when a larger model can plan the job and hand Ember small steps. Meta’s own model guide points difficult single-shot coding, knowledge-heavy questions and visual coding toward Muse Spark or Muse Wonder instead. No independent coding test is available for this article.

The useful scale marker is its 1,048,576-token context window, listed in Meta’s developer docs on November 16, 2026. That is a ceiling, not proof that a million-token prompt is necessary, inexpensive or more accurate. The practical question is whether your workflow can keep each assignment narrow, checkable and repeated often enough to justify an API model.

Meta's Muse Ember launch announcement on the Meta AI Research website.
Meta AI Research’s official announcement for Muse Ember.

Key takeaways

  • Model API ID: muse-ember-1.0; Meta lists Standard mode only.

  • The strongest fit is a small task delegated by a larger agent, such as a targeted edit, test run, classification or tool step.

  • Meta’s docs recommend Muse Spark or Muse Wonder for hard single-shot coding, knowledge-heavy answers and visual coding.

  • As of November 16, Meta lists Muse Ember in both Model API and Muse Code. Your account and client still need to expose the route.

  • API requests above 200K input tokens use the higher long-context rates in full. Check the pricing table before sending a large codebase or long history.

  • No Tabbit model integration or independent quality result was verified here.

Muse Ember at a glance

ItemMeta’s current documentationWhat it means for a decision
Model IDmuse-ember-1.0Set the exact ID in a compatible Meta API client
ModeStandard only on Model APIDo not expect a Contributor-mode variant
Input / outputText, image, video, audio and PDF in; text outIt can receive several media types, but that does not make visual coding its recommended use
Context / output limit1,048,576 input-context tokens; 524,288 max outputLarge limits are useful for selected workloads; long requests have different API rates
PositioningRepeated agent work, sub-agent steps, tool calls, routing and collectionsThink “many bounded tasks,” not “one model for every hard problem”
AccessModel API and Muse Code listed as availableConfirm account, billing, region and client route before committing a workflow

Source: Meta’s Muse Ember docs, checked November 16, 2026.

Choose Ember for repeatable subtasks, not every coding job

Meta describes Ember as an execution model for the work an agent repeats: coding subtasks, tool calls and routing. That suggests a practical pattern: a larger model plans, Ember performs a small step, and a test or reviewer checks the result. The pattern is most useful when you can say what a successful step returns before you run it.

WorkloadStarting choiceWhy
Apply a named, local code edit with a clear acceptance testEvaluate Muse Ember as a delegated stepThe task is bounded and repeatable; still review the diff and test output
Run the same extraction or classification across many recordsEvaluate Ember with a fixed schema and sample reviewMeta lists routing, extraction and work over large collections as target uses
Decide architecture from an ambiguous codebase or solve a hard one-off bugStart with Muse Spark or Muse Wonder, as Meta recommendsPlanning and open-ended reasoning dominate the task
Build or debug a visual interface from screenshotsStart with Muse Spark or Muse WonderMeta directs visual coding to the larger models
Give an agent broad write, deploy or account permissionsDo not decide from model family alonePermission scope, test coverage, rollback and human review matter more than a model label

These are workload-based starting points, not a claim that Ember will be faster, cheaper in practice or more accurate for your repository. Meta’s launch post reports favorable benchmark comparisons, but those are vendor-reported results under its cited setups. They are not a substitute for a task set from your own codebase.

A small evaluation you can run safely

Before using Ember for repeated coding work, pick a few representative tasks that already have a known expected result. Give each task a narrow instruction, a test command and a write boundary. Record the requested model ID, returned model ID, input length, tool calls, diff, test result, review changes and actual bill. Compare against your current model on the same tasks. Do not infer a general success rate from a handful of examples, and do not let a first run deploy or change production data.

The same discipline applies to a multi-model browser workflow. Keep public documentation and any signed-in workspace evidence separate, save URLs beside claims, and leave the output for review. Existing guides explain how Muse Spark 1.2 fits a signed-in browser workflow, how to connect an agent to an existing browser, and the first-task checklist. These guides do not establish that Muse Ember itself is integrated with Tabbit.

API access, Muse Code and the pricing threshold

Meta’s current docs list Muse Ember under its Model API as muse-ember-1.0 in Standard mode, and list Muse Code availability on pay-as-you-go and every plan. This is a documented rollout statement, not a guarantee that a particular account, country, provider or client has access. Check the model page, authentication guide and pricing table before you build against it.

Current Model API rates shown by Meta per one million tokens are:

Request sizeCached inputInputOutput
Up to 200K input tokens$0.004$0.10$0.50
More than 200K input tokens$0.012$0.30$1.50

The threshold applies to the request’s input length. Meta says a request above 200K is billed at the long-context rates in full, rather than charging only the tokens above 200K at the higher rate. That makes repeated agent history and large context bundles a cost decision. Trim irrelevant history, reuse cached prefixes where appropriate, and check the usage dashboard. The launch announcement on November 9 showed the standard input and output rates; the current docs also show cached-input and long-context rates. Recheck the live table before estimating a budget. In the Reddit launch thread, one commenter wrote, “For longer prompts, Muse Ember costs $0.30 per million input tokens.” Treat that as a cue to check the threshold, not as the source of current prices; Meta’s live table is authoritative.

Muse Ember accepts the same request body as Muse Spark according to Meta’s model docs, and the API supports Responses, Chat Completions and Messages endpoints. Tool calling, structured output, computer use, search grounding and media understanding are listed capabilities. A model capability does not automatically mean the client you use exposes every feature.

Where Tabbit fits, and where it does not

Tabbit Browser can help keep Meta’s model docs, API examples and your own implementation notes open while you investigate the route. It is a browser workspace around the research and review process. This article did not verify Muse Ember inside Tabbit’s model picker, a direct Tabbit-to-Meta integration or a completed Ember task. Choose and configure a supported model client separately.

If you are building a local browser agent, the Tabbit CLI complete guide and browser MCP vs CLI guide cover the browser connection layer. The Codex setup guide and Codex browser vs Tabbit CLI comparison explain client choices. For research, see web search vs browser control and the Muse Spark Open browser workflow. These are adjacent browser workflows, not model support claims.

Verdict: test Ember on repeatable, bounded work

Muse Ember is worth evaluating when your agent can hand it many small, verifiable jobs and you can inspect the outputs. It is a weaker starting point for a hard, one-shot coding question, visual implementation or open-ended planning because Meta’s own guide points those cases to Spark or Wonder. Check the model route and the long-context tier before wiring it into a high-volume workflow.

A useful next step is to choose a repeatable task, write its acceptance check first, then compare the same sample on the model route you already use and muse-ember-1.0. Keep the first run away from production writes, and decide from your own measured results rather than Meta’s benchmark headline.

FAQ

What is Muse Ember?

Muse Ember is Meta's hosted model for repeated agent work such as tool calls, routing, extraction and coding subtasks. Its Model API model ID is `muse-ember-1.0`.

Is Muse Ember good for coding?

Meta positions it for bounded coding subtasks inside an agent workflow, such as focused fixes and tests. Meta's docs direct hard single-shot coding and visual coding to Muse Spark or Muse Wonder.

How do I access Muse Ember?

Meta lists `muse-ember-1.0` on its Model API in Standard mode and lists Muse Code availability on pay-as-you-go and every plan. You still need to confirm your account, region, billing and client access.

Is Muse Ember available as open weights?

The Meta documentation checked for this article describes Muse Ember as an API-hosted model and does not provide a Muse Ember weight download. The available evidence supports the hosted route; it does not establish that no alternate distribution exists.

How large is Muse Ember's context window?

Meta lists a 1,048,576-token context window and a maximum output of 524,288 tokens. These are upper limits, not a recommendation to send every task the largest possible context.

Does long context change Muse Ember API pricing?

Yes. Meta's current table sets a higher rate when a request contains more than 200,000 input tokens, and the long-context rate applies to the whole request. Cached input, regular input and output have separate rates.

Take the next step

Let Tabbit work alongside you.

Research across tabs, automate repetitive browser work, and keep every piece of context within reach.