BRENNEN

Muse Ember

The jobs agents repeat

Meta's Muse Ember matched Claude Haiku 5.5 on an outside firm's tests at a third of the cost per task, though both list the same prices for text in and out.

Models · Checked 13 Nov 2026 · 11 sources · 9 min

Most offices have a job like this one. A folder fills up with invoices, and somebody opens each one, finds the total and types it into a spreadsheet. None of it is hard, and there is a great deal of it. Software can now be handed such work, and it goes about it the way the person did, one small step after another. A program that works through a job like this is called an agent.

An agent does not work out those steps by itself. At every step it sends a request to an AI model, a far larger program that usually runs on its maker's computers, and the model sends back the next move: which file to open, what to type. Every request is paid for, by length. Text is counted in tokens, a token being a short word or a piece of a longer one. The maker posts one rate for the tokens sent in, called input, and another for the tokens the model writes back, called output.

On 9 November 2026 Meta released Muse Ember, which it calls its smallest and most efficient model. Meta says it is designed for "the work agents repeat all day", and lists that work as tool calls, coding subtasks, routing, extraction and long document parsing. The invoice job is made of two of those. Asking another program to open a file is a tool call, and finding the total on a page is extraction.

Ember's rates start at $0.10 per million input tokens and $0.50 per million output tokens. Anthropic, a rival company, lists the same two rates for its Claude Haiku 5.5, a model Meta compares Ember with, on requests whose input is no longer than 100,000 tokens. Going by those rates, a job should cost the same on either model.

Artificial Analysis, an independent firm that tests AI models, found that it does not. It put both through its Intelligence Index, a fixed set of tests that produces one score, and recorded what the runs cost. Ember scored 36 at $0.14 per task. Haiku reached the same 36 at $0.41 per task. Both were run at their highest effort setting. Fourteen cents against forty-one is a small matter once. Suppose 1,000 tasks a day at those costs, though, and the two daily bills would be $140 and $410.

Cost of a task on the Intelligence Index, by Artificial Analysis

The same score of 36, at two costs

Both models were run at their highest effort setting.

Artificial Analysis

  1. Muse Ember
    $0.14 per task
  2. Claude Haiku 5.5
    $0.41 per task

Meta's summary is that Ember "matches Haiku's best score at about a third of the cost". Its own chart carries the figures before rounding: 35.8 for $0.136 on Ember's side and 36.2 for $0.411 on Haiku's. So the scores match only once each is rounded to 36, and the third holds, since $0.136 divided by $0.411 is 0.33.

A bill is a rate multiplied by a quantity. If the posted rates match and the bills do not, then either the amounts of text differ or some rate other than the two posted ones does. By Artificial Analysis's account it is both, and the invoice job shows where.

A hundred invoices, five million tokens

A model keeps nothing from one request to the next. Each request is answered from scratch, using only the text that arrived with it. So at every step an agent typically sends the whole record of the job so far: the instructions, every invoice already opened, every entry already made, and then the new invoice.

Suppose the folder holds 100 invoices, the agent deals with one at each step, and each step adds 1,000 tokens to the record. The first request carries 1,000 tokens, the second 2,000, and the hundredth 100,000. The average request, halfway between the first and the last, carries about 50,000, and there are a hundred of them, so the model is sent about 5 million tokens in all. Only 100,000 of them were new when they arrived. For each token the model meets for the first time, it is sent about fifty that it has been sent before.

Text a model has been sent before can cost its maker less to handle the second time, and both companies post a separate, much lower rate for input reused from an earlier request. The field calls this cached input. Meta's cached rate is $0.004 per million tokens, a twenty-fifth of its rate for fresh input. Artificial Analysis gives Haiku's as $0.01, a tenth.

Dollars per million tokens, as listed

The two price lists part only on cached input

Ember's rates start at these figures. Haiku's are for requests whose input is no longer than 100,000 tokens, and its cached rate is as Artificial Analysis gives it.

MetaAnthropicArtificial Analysis

Muse EmberClaude Haiku 5.5
Fresh input$0.10$0.10
Output$0.50$0.50
Cached input$0.004$0.01

Now price the invoice job, supposing every repeated token is billed as cached input. The 100,000 new tokens are a tenth of a million, so at $0.10 per million they cost 1 cent on either price list. The repeats, close to 5 million tokens, cost about 5 cents at Haiku's $0.01 per million and about 2 cents at Ember's $0.004. On both bills the rereading costs more than the reading, and it is charged at neither of the two rates the comparison began with.

Artificial Analysis gives this as the first of two reasons for the gap it measured. Cached input, the firm says, is most of the cost in the tests where it runs models as agents, and Ember's cached rate is 60% cheaper than Haiku's, $0.004 being 40% of $0.01. That cannot be the whole account. A 60% saving on one line of a bill takes less than 60% off the total, and the measured saving is about two-thirds.

The second reason is how much each model writes. Output carries the higher of the two posted rates: $0.50 per million is five times the $0.10 charged for input. And a model writes more than the short instruction that goes back to the agent. A model of this kind can work through a problem in writing before it answers. The field calls that working reasoning, and it is one of the costs Artificial Analysis counts. Meta's documentation describes a reasoning effort setting, chosen with each request, that runs from "minimal" to "max", the level used in the index runs. Artificial Analysis reports that Ember used roughly a third as many output tokens as Haiku.

In the invoice job, suppose one model writes 60,000 tokens over the 100 steps and the other does the same work in 20,000, a third as many. At $0.50 per million, that is 3 cents against 1 cent.

Add up the lines of the made-up job. At Haiku's rates, with the longer output, the bill is 1 cent for new text, 5 for rereading and 3 for writing, which makes 9 cents. At Ember's rates, with the shorter output, it is 1, 2 and 1, which makes 4. The two posted rates never differed, and one bill is less than half the other.

The job is invented, and so is its ratio. It shows only which lines of a bill can move while the two posted rates stay level. Artificial Analysis's cost per task is the same kind of sum done on real runs: a weighted average of token costs, counting fresh input, cached input, the storing of text in the cache, reasoning and answers. The runs are those of the Intelligence Index, which combines nine evaluations in four equally weighted categories: knowledge work, coding, science and general. The $0.14 and the $0.41 describe that one mix of tests, at maximum effort.

Five rows from outside, four from Meta

Meta has published no paper on Ember and no description of its design, its training data or its size. So "smallest" cannot be checked against a number. What can be checked is test results, and there it counts who ran each test.

A company testing its own product chooses the tasks, the instructions and the software around the model, and it knows its own model's habits better than a rival's. Meta's announcement carries a results table, and a four-page evaluation report names the party behind each number. Five rows are Artificial Analysis's, the Intelligence Index among them. For four others Meta ran the rival models itself, in runs it calls "best-effort", with prompts, tools and runtime that "may not be optimally tuned for proprietary models". No independent result for Ember on those four was found.

Even the outside score is an average, and two models can reach the same average by different routes. Artificial Analysis names knowledge as Ember's main weakness: on a test called AA-Omniscience, its accuracy was 21.1% against Haiku's 34.7%, mostly because it declined to answer 64% of the questions. Declining has a use. A model that answers when it does not know produces confident mistakes, which the field calls hallucinations, and the refusals held Ember's hallucination rate to 17.4%.

Accuracy on AA-Omniscience, by Artificial Analysis

Level on the index, apart on knowledge

Artificial Analysis

  1. Muse Emberit declined to answer 64% of the questions
    21.1%
  2. Claude Haiku 5.5
    34.7%

Meta's own documentation steers hard coding done in a single attempt, knowledge-heavy questions and visual coding to two other Meta models, Muse Spark and Muse Wonder. Its model-selection guide puts Muse Wonder on an agent's main loop, the part that runs the job as a whole, and Ember on sub-agents and background tasks, "which run many short, focused jobs where cost adds up". A sub-agent is a helper to which the main agent hands one piece of the job. On the index, Artificial Analysis places Ember 8 points behind Muse Wonder for about 1/50th of the cost per task.

One invoice, or a hundred and one

None of this fixes what a particular job will cost. With the two posted rates equal, Artificial Analysis's explanation puts the gap on two lines of the bill, the cached rate and the amount written. Suppose the folder holds a single invoice and the model writes only a few words. Nothing is reread, so almost the whole bill is fresh input, charged at $0.10 per million by both companies, and the two bills come out nearly equal. If a workload reuses little cached input or produces short outputs, its gap would be smaller than the index figure. A second outside tester, Vals AI, has priced Ember on its own index, at $0.153 per test, the lowest cost per test on that index.

A long job changes the bill in another way, because past a certain length the posted rates change. The hundredth request in the invoice job carries 100,000 tokens. A 101st would carry more, and above 100,000 tokens Haiku 5.5 moves to long-request rates of $0.50 input and $2.50 output per million. Meta's line is at 200K input tokens, K meaning thousand, and above it Meta bills three times the base rates on every token in the request. A record that grows with every step crosses Haiku's line first and Meta's later, and beyond the first line the two companies' posted rates no longer match.

Check it yourself

  1. Open Meta's announcement and find the cost-per-task chart to confirm 35.8 at $0.136 for Ember and 36.2 at $0.411 for Haiku, then read footnote 1 for the source of each row in the results table.
  2. Compare the Artificial Analysis page for Muse Ember with its page for Claude Haiku 5.5 to confirm that both show 36 and that the costs per task, $0.14 and $0.41, give "about a third" without Meta's chart.
  3. Compare Meta's pricing and rate limits page with Anthropic's pricing page to confirm the shared $0.10 and $0.50 rates on short requests, the different cached rates, and the length at which each company's higher rates begin.
  4. Read the Artificial Analysis launch analysis to confirm its two reasons for the cost gap, cached input and output tokens, and its figures for Ember on the knowledge test.
  5. Read the Artificial Analysis methodology to see the nine evaluations and four categories that make up the Intelligence Index.
  6. Read the Muse Ember evaluation report to see which party produced each number and the "best-effort" wording on Meta's runs of rival models.
  7. Open the Vals AI model page for a second outside measurement of Ember, including its cost per test.

Sources