What Is Muse Ember?

Muse Ember is Meta’s model for the steps an agent repeats, released on November 9, 2026. The model id is muse-ember-1.0. Meta’s research post names the job as tool calls, coding subtasks, routing, extraction, and long document parsing. The model page, read November 10, prices prompts up to 200,000 input tokens at $0.10 per million input tokens and $0.50 per million output tokens.

That sticker matches the short-prompt rates of Claude Haiku 5.5 and GPT-6 Luna. The published cache rates do not. Ember’s cached input is $0.004 per million tokens on that tier. Haiku’s cache-hit rate, for prompts up to 100,000 tokens, is $0.01. Luna’s cached input is $0.01. On the Artificial Analysis Intelligence Index in Meta’s post, Ember and Haiku both score 36 at max reasoning, and Luna scores 28. Haiku leads the same post’s FrontierCode row.

AI Agents Info read the research post, the model page, the API docs, the builder post, Anthropic’s pricing page, and the GPT-6 Luna model page on November 10, 2026. This page did not rerun the evaluations. The product listing is Muse Ember. How a harness wraps a model is What is an AI agent harness?.

What is Muse Ember?

Muse Ember is the worker model in the Muse family. The docs, read November 10, say to run Muse Wonder or Muse Spark on the main loop, and to hand bounded subtasks, tool calls, routing, and large batches of files to Ember. The research post uses the same split in one sentence: Muse Wonder plans, Muse Ember handles the volume.

The docs list a context window of 1,048,576 tokens and a maximum output of 524,288 tokens. The research post calls the output limit 512K, which is 512 × 1,024. Reasoning effort runs from minimal to max, and the default is max. The score table below is the max setting. Endpoints are Responses, Chat Completions, and Messages. Input can be text, image, video, audio, or PDF. Output is text. The docs also list computer use, tool calling, structured output, and search grounding.

On the API, Ember runs in Standard mode only. The docs say it has no Contributor-mode variant. It is available on pay-as-you-go and, the builder post says, in every Muse Code subscription. The research post says Muse Ember, Muse Spark, and Muse Wonder are also on Google Cloud and Oracle Cloud.

Claude Fable 5.2 came out the same day, for a different seat. That model is for long sessions, at $8 and $40 per million tokens. The write-up is What is Claude Fable 5.2?. Ember’s rivals on Meta’s own chart are Haiku 5.5 and Luna.

What does Muse Ember cost?

The model page prints two tiers. A request with more than 200,000 input tokens, including a long conversation history, pays the higher tier for that request. Muse Code subscriptions still allow Ember, and the docs say a request over 200,000 input tokens uses the plan’s allowance faster.

The short-tier input and output rates match two other models. The cache rate and the length where the price jumps do not.

Haiku’s cache column is the cache-hit price on Anthropic’s pricing page. Ember’s and Luna’s cache columns are the cached-input prices on their pages. A prompt of 150,000 tokens is still on Ember’s $0.10 tier and already on Haiku’s $0.50 tier. Luna’s model page says a prompt over 272K input tokens is priced at 2 times the input and cache rates and 1.5 times the output rate for the whole request, which is the $0.20, $0.02, and $0.75 row. Dividing Ember’s $0.004 by the $0.01 cache rate on the other two short tiers gives 0.4.

Meta also prints a cost per task, and the two posts do not use the same figure. The research post says Ember costs about $0.14 per Artificial Analysis task, and its comparison table lists $0.14 against $2.39 for Muse Spark 1.3 (max). Dividing $2.39 by $0.14 gives 17.1. The same table lists Spark at $1.25 input and $4.25 output per million tokens, and an Intelligence Index of 37 against Ember’s 36. The builder post says the estimated model cost at max effort is about $0.13 per Intelligence Index task. This page keeps both sentences. It does not average them.

Muse Code plans on the model page, read November 10, are $5, $15, $50, and $200 a month, named Everyday, High, Power, and Maximum. The Everyday bullet lists access to Muse Spark, Muse Wonder, and Muse Image. The API docs say Ember itself is included on pay-as-you-go and every plan. Those subscription prices are a seat. The token table above is the API bill.

The cost calculator prices Claude Haiku 5.5 and GPT-6 Luna from their provider pages. Ember’s rates are the table on this page. A separate check, for the tool list a harness sends on every turn, is the prompt checker. Why that JSON can cost more than the instructions is Why tool JSON costs more than your system prompt.

How do the agent scores compare?

Every cell below is max reasoning, copied from Meta’s November 9 research post. Artificial Analysis rows are marked AA. The other rows are Meta’s evaluation suite. Terminal-Bench 4.0 uses each model’s best published harness: mini-swe-agent for Muse Ember, Claude Code for Claude Haiku 5.5, and Codex for GPT-6 Luna. A higher score in that row is the model plus that harness.

The Coding Agent Index row is printed as 46, 37, and 31, without a percent sign. The Personal Superintelligence Index is an internal composite of eight evaluations: Agents' Last Exam, AutomationBench, FocusBench, JobBench, OfficeQA Pro, GDP.pdf, GDP.xlsx, and VBT. Those pieces overlap the rows above, so the index is a summary. Adding it to AutomationBench would count some of the same work twice.

Read as a buying table, Ember leads the agent-workflow and computer-use rows Meta chose to print, ties Haiku on the Intelligence Index, and trails Haiku on FrontierCode, the row the post describes as real engineering tasks judged on mergeability. Luna does not lead a row in this table. The research post’s sentence on cost is that Ember matches Haiku’s best score at about a third of the cost and beats every Luna setting. The post does not print the token counts behind “a third.”

Partner lines on the same posts are their tests. Cloudflare’s CTO says Ember ran up to 30% faster at the tail than the comparable model Cloudflare tested, and the sentence does not name that model. You.com’s inference lead says a p95 of 2.2 seconds, about 4 to 13 times faster than leading open-weight models, and 36% to 48% fewer output tokens. The sentence does not name those models. Box’s VP of AI products describes contract scoring and candidate screening on Box’s own eval. Treat those as vendor reports.

On safety, the research post says Meta evaluated Ember under its Superintelligence Scaling Framework, including chemical, biological, cybersecurity, and loss-of-control risks, and that the evaluations it ran found Ember below that framework’s high-risk thresholds. It also says Meta tested whether the model respects user intent, handles sensitive information, resists instructions hidden in pages and tool outputs, and asks a person before high-impact actions. Those are Meta’s results.

How do you put Muse Ember in a harness?

The builder post’s rule is a handoff, not a new framework. The planner gets the change that touches many files, the dependencies, and the acceptance checks. The Ember worker gets one brief: the files it may change, the commands that check the work, and the result to return. When the worker finds a dependency the brief left out, the finding goes back to the planner before the next edit. Orchestration is that choice of who goes next. Tool use is the command the worker actually runs.

The published coding example ports 16 Pygments lexers from Python to Rust. The checkout named in the post is a43b45dc. Muse Wonder plans and does not open the source itself. One worker reads the Python, the Rust engine, and the checker, and returns a map of at most 60 lines. Wonder writes a brief per lexer. Sixteen Ember agents then port in parallel, each in its own git worktree. The worker definition sets the model to muse-ember-1.0, effort to max, and a cap of 120 turns. The session allows 17 agents. The post says the 16 porters finished within about 6 minutes of each other.

The result the post reports: all 16 lexers matched Python on every one of 383 test inputs, and on 62 of 62 files from 18 public repositories that were supplied only after the code was frozen. The Rust commands ran 3.0 to 7.3 times faster than Python on about 1 MB of input. Ember handled 158 million input tokens. Wonder handled 4.4 million. The run cost about $3.96 at pay-as-you-go prices, $2.12 for Wonder and $1.84 for Ember. The post says results vary from run to run. AI Agents Info did not rerun it.

The same post’s photo job shows the other shape, one decision repeated thousands of times, and it prints two bills. One sentence says 10,000 photos were captioned, renamed, and sorted for $1.59, of which $0.50 was Ember. A later sentence says the job cost $1.33, $0.81 for Wonder and $0.52 for Ember, and that the same photo requests at Muse Wonder prices would have been about $20. The albums agreed with the dataset’s own labels for about 86% of photos. This page does not know which total is the one to quote, because the post prints both. The recipe’s worker uses effort low and a cap of 20 turns, and the pipeline sends eight photos per request at minimal reasoning. That is a different effort from the score table.

A direct API call returns text. The application, or Muse Code, runs tools and sends the results back. The docs’ first call is a POST to https://api.meta.ai/v1/responses with a bearer MODEL_API_KEY and "model": "muse-ember-1.0". The builder post says an existing OpenAI SDK client keeps its code, sets base_url to https://api.meta.ai/v1, and passes the key explicitly because the SDK does not read MODEL_API_KEY on its own. GPT-6 Luna’s model page, read the same day, lists text input and output, image input, and no audio or video. Ember’s docs list all four of those inputs, plus PDF.

The sandbox in the cookbook session is Muse Code’s, with network set to restricted in the migration command. A harness you already run can do the same split by choosing the model id per call. The comparison of shipping runtimes is Best AI agent harness. Coding products with their own seat prices are Best AI coding agents. The loop those products implement is How do AI agents work?.

What should a buyer check before switching a worker?

The default effort is max, and the benchmark table is max. A batch job can set a lower effort, as the photo recipe does, and this page has no score table for that setting. The price jump is at 200,000 input tokens for Ember, 100,000 for Haiku, and 272K for Luna. A harness that stuffs a whole repository into one prompt can leave Ember’s cheap tier without leaving Luna’s.

Terminal-Bench is three harnesses, not one. FrontierCode is the row Haiku wins. Partner latency numbers do not name the other model. The two cost-per-task sentences, $0.14 and $0.13, are both in Meta’s posts from the launch, and the photo example prints $1.59 and $1.33. Use the token table for a bill you can recompute, and use the score table for the jobs Meta actually published.

Claude is the product that serves Haiku. The free agent picker is a directory shortlist, separate from this model choice. List a harness you ship at Submit.