The short version: Meta Muse Ember, released November 9, 2026, is a small, fast AI model priced at $0.10 per million input tokens and $0.50 per million output tokens on Meta’s own model page (checked November 10, 2026). That is one fortieth of what Meta charges for its big Muse Wonder model. Meta built Ember to do the volume while a bigger model does the planning. The surprise is in Meta’s own demonstration: in a job where Ember captioned and sorted 10,000 photos, the planner still took about 61 percent of the $1.33 bill. The cheap part of an AI job is now the doing. What decides the cost and the quality is the work order the planner writes, and writing a good work order is a skill every owner who runs a crew already has.
Meta’s launch post for Muse Ember publishes the brief it hands each worker: the task, the relevant context, the files the worker may change, the checks it has to pass, and the result it should bring back. Strip out the software vocabulary and that is a job ticket. This piece explains how Meta’s split between planner and worker actually runs, does the division on Meta’s own example jobs, checks two claims circulating about the model, and turns the brief into three habits a small business can use with any AI tool.
What is Meta Muse Ember?
Muse Ember is the smallest model in Meta’s Muse family, which also includes Muse Spark and the large Muse Wonder. According to Meta’s research post, it is the first model Meta trained using Muse Wonder Full, its biggest internal model. It reads up to 1 million tokens at once, can write up to 512,000 tokens in a reply, and Meta says it handles text, images, audio and video.
You reach it three ways: through Meta’s developer API under the name muse-ember-1.0, inside Muse Code, Meta’s coding agent, where Meta’s model page lists subscription plans starting at $5 a month, and through Google Cloud and Oracle Cloud. None of the pages we read on November 10 say the weights are downloadable or name a license, so treat it as a rented model, not one you can run on your own machine. That is a real difference from Muse Spark Open, which Meta did release for download.
The full price list, read off Meta’s page on November 10:
Requests up to 200,000 input tokens: $0.10 per million input tokens, $0.50 per million output tokens, $0.004 per million for cached input.
Requests over 200,000 input tokens: $0.30 input, $1.50 output, $0.012 cached. Three times the rate.
Muse Wonder, for comparison: $4.00 input, $20.00 output, $0.20 cached, on Meta’s Wonder page the same day.
How does the planner and worker split work?
The idea is older than AI. One experienced person looks at the job, decides how it should be done, writes it up, and hands the repetitive part to a crew. Meta’s launch post describes the software version plainly: “Your application or agent harness chooses the model for each call.” In other words, nothing inside Ember decides when to call the big model. The tool around it does, and it carries the notes from one step to the next.
Meta’s photo example shows the whole loop. The instruction was one line: caption, rename and sort every photo in an inbox folder into albums. Muse Wonder captioned a small sample to decide what the albums should be, wrote one script to do the sorting and tested it, then started 16 agents and gave each a share of the 10,000 photos. Each agent ran Wonder’s script, which sent photos to Ember eight at a time, shrunk to 512 pixels, on its lowest reasoning setting. Ember sent back a caption and an album for each one. At the end, Wonder checked that every photo had been filed. Meta reports eight albums and no failed requests.
The second example, converting 16 pieces of a Python program to a faster language, Rust, shows the same shape at a bigger scale. Wonder never read the original code itself. One agent mapped it, Wonder turned that map into a separate brief for each of 16 Ember workers, and each worker read its piece, rewrote it, tested it against the original and committed only if it passed. Meta says all 16 matched the original on 383 test inputs and on 62 real files.
Two details in Meta’s worker instructions are worth noticing. A worker that finds something outside its brief, like a file it does not own or a requirement that is unclear, is told to send the finding back to the planner before continuing. And a worker that hits the same failure three times is told to stop and report itself blocked rather than keep trying.
Who pays for what? The division Meta did not do
Meta gives the cost of both example jobs but not the shares, so here they are.
The photo job: $1.33 in total, $0.81 for Wonder and $0.52 for Ember. Ember handled every one of the 10,000 photos, and the planner still took 61 percent of the bill. The launch post also gives a second set of figures for the same job, $1.59 in total with $0.50 for Ember, which would put the planner at 69 percent. We could not tell from the page which figure is final, but both point the same way. Per photo, the job cost about one hundredth of a cent. Meta says the same photo requests sent to Wonder alone would have cost about $20.
The code job: $3.96 in total, $2.12 for Wonder and $1.84 for Ember. Ember read 158 million input tokens and Wonder read 4.4 million, so Ember did 97 percent of the reading, and Wonder still took 54 percent of the money.
That is not a flaw in the setup. It is the setup working. When one model costs forty times as much per token as the other, even a short stretch of planning outweighs a long stretch of doing. So the bill is no longer driven by how much work there is. It is driven by how much thinking the plan needs and whether the plan is right the first time. A brief that is wrong is not wrong once; it is copied to all 16 workers, and you pay to do the job twice.
For an owner, that is good news with a lesson in it. The repetitive part of office work is getting close to free. The part that still costs money, and still decides whether the result is any good, is the part that looks most like what an owner already does all day: deciding what the job is, what counts as done, and what to check.
Is Muse Ember cheaper than Claude Haiku 5.5?
Per token, no. Meta’s chart compares Ember with Anthropic’s Claude Haiku 5.5, and the two carry the same list price: $0.10 in and $0.50 out, by Anthropic’s pricing page as we reported on October 7. Meta’s claim that Ember matches Haiku’s best score “at about a third of the cost” is measured per benchmark task at the highest reasoning setting, which means Ember used fewer tokens to reach the same score in Meta’s test. That is a meaningful claim, but it is Meta’s own measurement, and it depends on how much a model thinks rather than on the price list.
Where the two genuinely differ is on long requests. Haiku 5.5 bills a whole request at five times the rate once the prompt passes 100,000 tokens. Ember waits until 200,000 tokens and then charges three times. Take a 150,000-token prompt, roughly a long handbook plus a question, with a 1,000-token answer. On Haiku 5.5 that is about 7.8 cents. On Ember it is about 1.6 cents. The two models count words into tokens differently, so the same document will not land on exactly the same number, but the gap at that length is real.
Two claims worth checking
“Smarter than Haiku.” TechCrunch summarized Meta’s chart as showing Ember cheaper and smarter than competitors including Haiku. Meta’s own table is more modest. On the Artificial Analysis Index, the overall score Meta leans on, Ember and Haiku 5.5 both score 36. Ember wins clearly on AutomationBench, an office-workflow test, at 37.8 to 13.9, and loses on the FrontierCode 2.0 coding test, 27.2 to 34.9. A footnote in Meta’s research post adds that the Terminal-Bench comparison ran each model inside a different software harness. Read it as a strong small model that trades blows with Haiku, not one that beats it everywhere.
Audio and video. Meta’s pages say Ember natively understands text, images, audio and video. Google Cloud’s listing for the model, marked Preview, says it accepts text and image inputs only. If you plan to use it on voicemails or job-site video through Google Cloud, confirm that first.
What does Muse Ember change for a small business?
Most owners will never call Muse Ember by name. They will meet it, or something like it, inside the tools they rent: the inbox sorter, the photo organizer, the bookkeeping assistant that tags receipts. Meta is explicit about splitting work between a planner and a cheap worker, and Anthropic pitches Haiku 5.5 at the same kind of high-volume sorting and extracting. The pattern is becoming normal.
Picture a remodeling contractor with years of job-site photos spread across three phones: before shots, rough-in, finished rooms, warranty callbacks. Sorting them by job and stage is exactly the kind of bounded, repetitive work Ember was built for, and at Meta’s figures 10,000 photos is a little over a dollar of model time. But Meta also reports that Ember’s albums agreed with the dataset’s own labels on about 86 percent of photos. On 10,000 photos that leaves roughly 1,400 in a different album than a person would have chosen, and nothing tells you which ones. The AI does the hours of sorting; a person still walks through each album before anything goes to a customer. That is time handed back, not judgment handed over.
Three habits from Meta’s own playbook
Write the brief like a work order. Meta’s template works for any AI tool, including a chatbot on your phone: what the task is, what it needs to know, what it may and may not touch, how you will check the result, and what you want back. The last two are the ones owners skip. When we looked at Mistral Large 4’s office-work scores, the best AI did most of the work and still lost points to one broken business rule. A rule that is written down is a rule the worker can follow.
Ask your vendor who plans and who checks. If a tool you pay for runs on a planner and a worker, ask which model writes the plan, which does the volume, and what check runs at the end. A tool that can answer all three is one you can trust with more. A tool that cannot is guessing on your behalf.
Keep a stop rule. Meta’s workers stop after three identical failures and report back. Ask whether yours do, or whether they quietly keep going. The Arena Alignment Index found AI agents reporting work as finished when it was not, and a stop rule is the cheapest guard against that. It is also the instruction a good foreman gives a new hire on day one.
Frequently Asked Questions
What is Meta Muse Ember?
Muse Ember is Meta’s smallest and cheapest Muse model, released November 9, 2026, built for high-volume repetitive work such as sorting, tagging, extracting and small coding tasks. Meta designed it to work under a bigger planning model like Muse Wonder, which writes the instructions while Ember does the volume.
How much does Muse Ember cost?
Meta’s model page, checked November 10, 2026, lists $0.10 per million input tokens and $0.50 per million output tokens for requests up to 200,000 input tokens, and $0.30 and $1.50 above that. Cached input costs $0.004 per million. Muse Code subscription plans that include it start at $5 a month.
Is Muse Ember open source?
None of Meta’s pages we read on November 10, 2026 describe Muse Ember as open weights or name a license. It is available through Meta’s API, Muse Code, Google Cloud and Oracle Cloud, so treat it as a rented model rather than one you can download and run yourself.
Should a small business choose Muse Ember or Claude Haiku 5.5?
Both list at $0.10 in and $0.50 out per million tokens, and both score 36 on the Artificial Analysis Index in Meta’s table. Ember is cheaper on prompts over 100,000 tokens and wins Meta’s office-workflow test; Haiku wins Meta’s coding test. Test both on your own work before choosing.
If you could hand one repetitive job in your business to a crew of cheap AI workers tomorrow, what would the work order say? Tell us in the comments.
