Meta releases Muse Ember at Claude Haiku’s price. Where it runs, and when to use it

What Meta released

On November 9, 2026, Meta introduced Muse Ember, “our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.” Meta says it’s the first of its specialized models “trained using Muse Wonder Full,” and that it rounds out the lineup: “Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable coding and agentic tasks at scale.”

Its ID is muse-ember-1.0. The docs page gives it a context window of 1,048,576 tokens, up to 524,288 output tokens, and text, image, video, audio and PDF input. Reasoning effort runs from minimal to max, and “the default is max,” so a request that sets no effort gets the deepest and most expensive setting.

It isn’t open-weight. Meta’s models page lists Ember among the models “hosted on Meta Model API,” and names Muse Glimmer and Muse Spark Open as the ones you run yourself.

Where you can use it

• Muse Code. “Muse Ember is on Model API and in Muse Code, on pay-as-you-go and with any subscription,” says its docs page. The subscription guide lists it from the Everyday Usage plan up, which the Muse Code page prices at $5 a month.
• Meta’s Model API, as muse-ember-1.0, on the Responses, Chat Completions and Messages endpoints.
• Google Cloud and Oracle Cloud. The announcement says “Muse Ember, Muse Spark and Muse Wonder are now available on Google Cloud and Oracle Cloud,” and Google’s model garden lists Ember.
• OpenRouter and OpenCode. OpenRouter has it as meta/muse-ember, and OpenCode’s Console as Muse Ember 1.0, both at Meta’s price.
• Not Cursor or GitHub Copilot, so far. Cursor’s model list has Muse Wonder but not Ember, and Copilot’s has no Muse model.

What it costs

Prices per million tokens, from each company’s own page on November 9, 2026:

Model	Input	Cached input	Output	Long prompts, for the whole request
Muse Ember	$0.10	$0.004	$0.50	Over 200K input tokens: $0.30, $0.012 and $1.50
Claude Haiku 5.5	$0.10	$0.01	$0.50	Over 100,000 tokens: $0.50, $0.05 and $2.50
GPT-6 Luna	$0.10	$0.01	$0.50	Over 272K input tokens: twice the input and cached rates, 1.5 times the output
Muse Spark 1.3	$1.25	$0.15	$4.25	Same rate at any length
Muse Wonder	$4.00	$0.20	$20.00	Same rate at any length

Meta’s rates are on its pricing page, Anthropic’s on its own, and OpenAI’s on its pricing and Luna pages.

So Ember isn’t cheaper per token than its two rivals. TechCrunch wrote: “At 10 cents per million input tokens, Ember is significantly cheaper than frontier competitors like Haiku and Luna.” All three list the same $0.10 and $0.50. The difference is in the bill for a finished task. Artificial Analysis says Ember’s “cached input, most of the cost in our agentic evaluations, is 60% cheaper,” and that it “uses roughly a third as many output tokens as Claude Haiku 5.5.”

Four things to know before you budget:

• Long requests cost three times as much. A request with more than 200K input tokens “is billed at the long-context rates in full, not only for the tokens past 200K.” The count includes cached tokens and, when you continue a stored conversation, that conversation.
• Reasoning is billed as output. Meta’s reasoning guide says reasoning tokens count “toward billed output tokens,” and Ember reasons at max unless you set a lower effort.
• There’s no batch price. Anthropic and OpenAI list batch processing at half price, $0.05 in and $0.25 out for Haiku 5.5 and Luna. Meta’s pricing page lists no batch rate.
• On the API, your content isn’t used for training. Ember “runs in Standard mode only,” where “your content is not used to train Meta models.” A Muse Code subscription is different: “Muse Code subscriptions default to Contributor mode,” which lets Meta train on your content until you turn it off in /settings.

How good it is

The independent numbers are Artificial Analysis’s, from its own article and the results Meta’s announcement reproduces, all at max effort:

Test	Muse Ember	Claude Haiku 5.5	GPT-6 Luna
Intelligence Index	36	36	28
Cost per index task	$0.14	$0.41	$0.15
Coding Agent Index, each in its own lab’s agent	46	37	31
FrontierCode 2.0	27.2%	34.9%	25.2%

Ember ties Haiku 5.5 on the index at about a third of the cost, and its model page there measures 449 output tokens a second. It sits “8 points behind Muse Wonder for about 1/50th of the Cost per Task.” By Meta’s own table, Muse Spark 1.3 scores one point more, 37, at $2.39 a task.

It has two weak spots. Coding is “mixed”: Ember beats Haiku 5.5 on Terminal-Bench 4.0, 40.9% to 32.8% in Artificial Analysis’s run, but trails it on FrontierCode, which judges whether a maintainer would merge the change. And facts: “Knowledge is the main weakness, but it rarely hallucinates.” It answers 21.1% of AA-Omniscience questions correctly, “mostly because it declines to answer 64% of questions.” That keeps its hallucination rate at 17.4%, “half of Claude Haiku 5.5’s.” Give it the documents or a search tool before you ask it for facts.

Meta’s own table is kinder to Haiku in one row and harder to check in others. It gives Haiku 39.2% on Terminal-Bench, Anthropic’s own Claude Code figure, and its evaluation report says the rival runs Meta did itself “remain best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models.” Those include the computer-use test, OSWorld 2.1, where Meta puts Ember at 69.5% with partial credit, against 68.9% for Haiku 5.5 and 50.5% for Luna.

How to use it in Muse Code

To run a whole session on Ember, start with muse --model muse-ember-1.0, or switch with /models in a running session, as the configuration guide describes.

The use Meta built it for is different: Muse Wonder or Muse Spark leads, and Ember runs the subagents. Muse Code doesn’t do that by itself. “By default, a subagent runs on the same model as your main session,” says the custom agents guide, and the lead “doesn’t switch its subagents to a lower-cost model on its own.” You have two ways to ask:

1. Name the model in your prompt, for a single request. Meta’s example: “Split the test suite into four parts and run each part in a subagent using muse-ember-1.0. If that model isn’t available, stop and tell me.” Use the exact ID: “With only the model’s name, the lead guesses an ID, and Muse Code rejects a wrong one.” Meta adds that this works “when workflows are available in your build.”
2. Define an agent, for repeated runs. Save this as agents.json and start Muse Code with muse --model muse-wonder-1.0 --reasoning-effort high --agents "$(cat agents.json)":

{
  "subagent": {
    "description": "Executes one well-defined coding part of a larger plan, such as implementing a module, adding tests or fixing a bug.",
    "prompt": "You receive one part of a larger plan from the lead. Complete that part as the task describes, creating or changing the files it needs. Run the checks the task names. Reply with what you changed and whether the checks pass.",
    "model": "muse-ember-1.0",
    "effort": "medium"
  }
}

Muse Code doesn’t save what you pass with --agents, so pass it each time, or keep the definition as a Markdown file in .agents/agents/ in your project. The built-in general-purpose subagent still runs on the lead’s model unless you define it too, so “a part meant for a Muse Ember subagent runs on Muse Wonder instead and costs more.” Before a long run, check that muse-ember-1.0 is in your /models list. On a subscription, a request over 200K input tokens “uses your plan’s usage faster.”

How to use it in Claude Code and other agents

Meta’s coding agents guide connects Claude Code to its API with a Meta API key, here in MODEL_API_KEY. You pay Meta per token, and your Claude plan isn’t used:

export ANTHROPIC_BASE_URL="https://api.meta.ai"
export ANTHROPIC_AUTH_TOKEN="$MODEL_API_KEY"
export ANTHROPIC_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_OPUS_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_SONNET_MODEL="muse-spark-1.3"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="muse-ember-1.0"
export CLAUDE_CODE_SUBAGENT_MODEL="muse-ember-1.0"
export ENABLE_TOOL_SEARCH="true"

The guide’s advice is “Pin every model alias,” so that no path “falls back to a Claude model Model API doesn’t serve,” and to use ANTHROPIC_AUTH_TOKEN, not ANTHROPIC_API_KEY. Our guide to other models in Claude Code covers the other routes.

For Codex, Meta says to set muse-wonder-1.0 as the default model and switch to muse-spark-1.3 “for faster, lower-cost sessions.” It doesn’t suggest Ember there. In your own agent loop, Ember “accepts the same request body as Muse Spark, so switching most examples is a model ID change.”

When to pick Ember, and when not to

Meta’s Choose a model page sorts the work:

1. Subagent tasks, tool loops, routing, classification, and the same job over thousands of files: Ember.
2. The main loop of a coding agent, and changes across many files: Muse Wonder, or Muse Spark when the main loop has to cost less.
3. Hard single-shot coding problems, knowledge-heavy questions and visual coding: “use Muse Spark or Muse Wonder rather than Muse Ember.”
4. Audio. Ember accepts it, but Meta’s table sends audio understanding to Muse Spark 1.2.
5. Large offline jobs where half price matters more than speed: Haiku 5.5 or Luna through their batch APIs.
6. A model on your own hardware: not Ember. See running a coding agent on a local model.

Ember in Enjoy

Enjoy runs Muse Code on your Muse Code subscription, and its Model column lists the models Muse Code offers your account. Muse Code only updates that list when its own app starts, so a model released today can be missing: in the desktop app, select the circular arrow beside Model and Enjoy has Muse Code fetch its latest models.

A conversation can use a different model for each message, so Muse Wonder can plan a change and Ember can do the routine steps in the same thread, and you can hand the conversation to Claude Code or Codex when you want a second opinion. Enjoy shows what’s left of each Muse Code limit once it has replied in the project.

Enjoy doesn’t add Meta’s models to another lab’s account, and Contributor mode stays on in Enjoy until you turn it off in Muse Code’s settings. Enjoy is a monthly subscription, with the first seven days free (pricing).

What Meta hasn’t said

• How big Ember is. It’s “our smallest,” with no parameter count.
• Whether weights will follow. The announcement says only: “We have even more efficient models in training and look forward to sharing what’s next.”
• How much of a plan it uses in Muse Code compared with Muse Spark or Muse Wonder.
• Which Muse Code version first offers it. The newest entry in the changelog, 1.4.13, doesn’t mention Ember.
• A knowledge cutoff. The announcement, docs page and model page don’t give one.

We’ll update this memo when Meta fills these in.