Muse Ember tested: Haiku’s price, a third of the cost per task

The answer in brief

Muse Ember is the smallest and cheapest model in Meta’s Muse family. Meta Superintelligence Labs released it on 9 November 2026 to do the repetitive steps inside AI agents: tool calls, coding subtasks, routing, data extraction and long-document parsing.

• Price: $0.10 per million input tokens and $0.50 per million output tokens on the Meta Model API. That is the same list price as Claude Haiku 5.5 and GPT-6 Luna. Once a request passes 200,000 input tokens, all of it is billed at three times the rate.
• Cost per job: on version 5.0.1 of the Artificial Analysis Intelligence Index, Ember at max effort scores 36 for $0.14 a task. Haiku 5.5 reaches the same 36 for $0.41. GPT-6 Luna peaks at 28.
• Speed: Ember takes 3.1 minutes per index task on Artificial Analysis’s measurements. Haiku 5.5 at max takes 14.7.
• Our test: on four everyday agency jobs, Ember at high effort matched GPT-6.1 Terra’s accuracy for about half Terra’s cost, and it never misread a figure on a badly compressed screenshot. GPT-6 Luna at medium was at least as accurate on search terms and cost less, because Ember thought far longer at the same effort setting.
• Weak spots: it trails Haiku 5.5 on FrontierCode 2.0, a test of mergeable fixes in real repositories, and Meta sends hard coding problems to its bigger models. Four of Meta’s ten headline comparisons come from its own internal tests. At max effort in our test, it once returned an empty answer.
• Access: the Meta Model API and Muse Code, as muse-ember-1.0, plus OpenRouter and Google Cloud. There are no downloadable weights. The UK is not on Meta’s list of restricted territories.
• Our view: if you run Haiku 5.5 or Luna at high effort inside agent loops, Ember is the model to test next, at high effort rather than the default of max. For short single-step jobs, compare it with Luna at medium before you switch.

What is Meta Muse Ember?

Muse Ember is a small reasoning model that Meta built to work underneath a bigger one. In a typical agent, a large model plans the job, splits it into pieces and checks the result. Ember is meant to do the pieces, many at a time and cheaply. Meta’s launch post sums up the split in one line: “Muse Wonder plans, Muse Ember handles the volume.”

Ember joins the models Meta Superintelligence Labs has shipped this year: Muse Spark in April, the open-weight Muse Glimmer in August and the frontier-class Muse Wonder in October. Meta says Ember is the first model it trained using Muse Wonder Full, the stronger Wonder variant that runs only inside Muse Code. Our guide to the Muse family explains how the models and the Muse agent app fit together.

Meta has not disclosed Ember’s size or parameter count, and the weights stay private. Despite the “smallest” label, you cannot run it on your own hardware. Meta’s open-weight options are Muse Glimmer and Muse Spark Open.

The developer docs list the specifications:

• a context window of 1,048,576 tokens and up to 524,288 output tokens;
• input of text, images, video, audio and PDF, with text output;
• six reasoning effort levels, from minimal to max, with max as the default;
• the Responses, Chat Completions and Messages endpoints, with tool calling, structured output, search grounding and computer use.

On safety, Meta says it tested Ember under its Meta Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks, and that it stayed below the framework’s high-risk thresholds. Meta also says it checked how Ember behaves when acting for a user, including whether it resists instructions hidden in web pages, documents or tool outputs. The launch post gives no scores for those agent tests.

How much does Muse Ember cost?

On prompts up to 200,000 input tokens, Muse Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. Above 200,000 input tokens the rates triple to $0.30, $0.012 and $1.50, and Meta then bills every token in the request at the higher rates. These are US dollar list prices; Meta’s pricing page quotes them without mentioning VAT.

On short prompts, Ember, Claude Haiku 5.5 and GPT-6 Luna cost exactly the same. TechCrunch’s launch report called Ember significantly cheaper than both at 10 cents per million input tokens, but Anthropic and OpenAI list the same 10 cents. Ember’s price advantage shows up per task, which the next section covers.

If the table extends beyond the screen, scroll sideways to view all columns.

Model	Input	Cached input	Output	Higher rate starts at	Input, cached, output above it
Muse Ember	Input$0.10	Cached input$0.004	Output$0.50	Higher rate starts at200,000 input tokens	Input, cached, output above it$0.30, $0.012, $1.50
Claude Haiku 5.5	Input$0.10	Cached input$0.01	Output$0.50	Higher rate starts at100,000 input tokens	Input, cached, output above it$0.50, $0.05, $2.50
GPT-6 Luna	Input$0.10	Cached input$0.01	Output$0.50	Higher rate starts at272,000 input tokens	Input, cached, output above it$0.20, $0.02, $0.75
Muse Spark 1.3	Input$1.25	Cached input$0.15	Output$4.25	Higher rate starts atNo higher rate	Input, cached, output above itNo higher rate

Two details separate the three small models. Ember’s cached input costs $0.004, 40% of the rate Haiku and Luna charge, which helps agents that resend the same instructions and files on every call. On long prompts, Ember sits in the middle: Haiku’s price jumps five times at 100,000 tokens, Ember’s triples at 200,000, and Luna’s input price only doubles at 272,000.

The 200,000-token line is easy to cross by accident. When a request continues a stored conversation with previous_response_id, Meta counts that conversation as input, so a long agent session can tip every later call into the higher tier. On a Muse Code subscription, those requests use up your plan faster.

Web search grounding adds $2.50 per 1,000 searches. Rate limits are 3,000 requests and 4 million tokens a minute, shared by every key in a team. Ember is also included in every Muse Code subscription, which runs from $5 to $200 a month.

Is Muse Ember cheaper than Claude Haiku 5.5 and GPT-6 Luna per job?

Yes, at the settings where the three models do their best work. Artificial Analysis runs every model through the same nine evaluations and records what each one costs per task. At max effort, Ember matches Haiku 5.5’s score of 36 for a third of the cost, and it beats every GPT-6 Luna setting.

If the table extends beyond the screen, scroll sideways to view all columns.

Model (effort)	Index score	Cost per task	Minutes per task	Seconds to first answer token
GPT-6.1 Terra (medium)	Index score38	Cost per task$0.17	Minutes per task2.7	Seconds to first answer token4.73
Muse Ember (max)	Index score36	Cost per task$0.14	Minutes per task3.1	Seconds to first answer token5.00
Claude Haiku 5.5 (max)	Index score36	Cost per task$0.41	Minutes per task14.7	Seconds to first answer token182.64
Claude Haiku 5.5 (xhigh)	Index score35	Cost per task$0.25	Minutes per task9.6	Seconds to first answer token51.96
Muse Ember (xhigh)	Index score34	Cost per task$0.12	Minutes per task2.7	Seconds to first answer token3.91
DeepSeek V4.1 Flash 1017 (max)	Index score33	Cost per task$0.60	Minutes per task15.2	Seconds to first answer token47.81
Claude Haiku 5.5 (high)	Index score30	Cost per task$0.16	Minutes per task6.4	Seconds to first answer token22.65
GPT-6 Luna (max)	Index score28	Cost per task$0.15	Minutes per task11.6	Seconds to first answer token77.37
GPT-6 Luna (xhigh)	Index score23	Cost per task$0.10	Minutes per task7.4	Seconds to first answer token19.85

The gap in waiting time matters as much as the gap in cost. An agent that calls a worker model hundreds of times a session waits for every first token. At max effort, Haiku 5.5 took more than three minutes to start answering on Artificial Analysis’s runs; Ember took about five seconds.

Ember does not get there by writing less. Artificial Analysis calls it somewhat verbose: it produced 115 million output tokens across the index, against a median of 107 million for comparable models. So the lower cost per task does not come from shorter answers, and neither Meta nor Artificial Analysis says where it does come from. The saving also depends on the job: on our own short, single-request tasks below, Ember cost more than Luna at the same effort setting.

Meta’s launch chart compared Ember only with Haiku 5.5 and Luna. OpenAI’s GPT-6.1 Terra, released on 26 October 2026, scores 38 at medium effort for $0.17 a task: two points more for three cents more, at a similar speed, even though Terra’s token prices of $1 and $5 per million are ten times Ember’s. Ember still sits on the efficiency frontier of the whole leaderboard, as Meta claims, because no model scores above 36 for less than $0.14. Teams choosing a small model should still test Terra alongside it; our GPT-6.1 Terra guide covers it.

How does Muse Ember score on Meta’s own benchmarks?

In Meta’s launch table, Ember leads or ties on nine of ten rows against Claude Haiku 5.5 and GPT-6 Luna, all three at max effort. The one loss is FrontierCode 2.0, where Haiku leads by 7.7 points. Who ran each test matters, so the last column says.

If the table extends beyond the screen, scroll sideways to view all columns.

Benchmark (what it tests)	Muse Ember	Claude Haiku 5.5	GPT-6 Luna	Who ran it
AutomationBench-AA (business workflows across SaaS apps)	Muse Ember37.8	Claude Haiku 5.513.9	GPT-6 Luna20.2	Who ran itArtificial Analysis
Agents’ Last Exam (end-to-end agent tasks)	Muse Ember61.6	Claude Haiku 5.557.0	GPT-6 Luna57.1	Who ran itMeta
Personal Superintelligence Index (Meta’s composite of eight tests)	Muse Ember46.8	Claude Haiku 5.536.1	GPT-6 Luna37.0	Who ran itMeta
Terminal-Bench 4.0 (coding in a terminal)	Muse Ember40.9	Claude Haiku 5.539.2	GPT-6 Luna16.4	Who ran itMixed: see below
Coding Agent Index-AA (coding inside each maker’s agent)	Muse Ember46	Claude Haiku 5.537	GPT-6 Luna31	Who ran itArtificial Analysis
FrontierCode 2.0-AA (fixes a maintainer would merge)	Muse Ember27.2	Claude Haiku 5.534.9	GPT-6 Luna25.2	Who ran itArtificial Analysis
GDP.pdf-AA (professional work from long PDFs)	Muse Ember22.8	Claude Haiku 5.520.8	GPT-6 Luna22.8	Who ran itArtificial Analysis
OSWorld 2.1 (computer use, partial credit)	Muse Ember69.5	Claude Haiku 5.568.9	GPT-6 Luna50.5	Who ran itMeta
PerceptionBench, with tools (fine visual detail)	Muse Ember80.6	Claude Haiku 5.578.6	GPT-6 Luna62.8	Who ran itMeta
Artificial Analysis Intelligence Index	Muse Ember36	Claude Haiku 5.536	GPT-6 Luna28	Who ran itArtificial Analysis

The Artificial Analysis rows are the most useful, because an outside lab ran every model the same way. The clearest win there is AutomationBench, where Ember completed 37.8% of Zapier’s business workflows in full against 13.9% for Haiku. That figure is the strict measure: every objective met and no rule broken. Artificial Analysis’s headline score for the same benchmark gives partial credit, so check which measure any AutomationBench figure you see quoted for Ember uses.

Four rows are Meta’s own runs, and Meta’s evaluation report says its runs of rival models are best effort, with prompts, tools and runtime that may not suit them. The Personal Superintelligence Index is an internal composite, and Meta dropped five of Agents’ Last Exam’s 105 tasks for every model because of storage limits or scripts that returned no score. OSWorld here gives partial credit; on the strict pass rate Ember scores 33.8% against 27.8% for Haiku.

The Terminal-Bench row compares different set-ups. Ember’s 40.9% is Artificial Analysis’s run in a simple open-source agent, mini-swe-agent. Haiku’s 39.2% is Anthropic’s own figure in Claude Code, and Luna’s 16.4% comes from the Terminal-Bench leaderboard in Codex. When Vals AI ran Ember on the same benchmark, it scored 39.39%.

The FrontierCode loss fits Meta’s own advice. Its docs tell developers to use Muse Spark or Muse Wonder for hard single-shot coding problems, knowledge-heavy questions and visual coding.

What do independent testers say about Muse Ember?

The outside results so far back Meta’s main claim, that Ember is unusually capable for its price, and add two warnings: it is wordy, and its reasoning settings behave unevenly.

Artificial Analysis ranks Ember at max effort second of 36 models in its price class for intelligence, against a class median score of 24. It rates Ember notably quick, at 3.1 minutes per index task against a median of 7.5, with 4.93 seconds to the first answer token and 451 tokens a second on Meta’s API.

Vals AI, which tests models on finance, legal, tax and coding work, puts Ember 14th of 37 models on its Vals Index, at 60.48%. When it published its launch results, Vals said Ember’s cost of $0.15 per test was the lowest of any model on the index. Ember did best on legal work: fourth of 38 on the Vals Legal Index (75.69%) and second of 38 on HLAB-Vals-1.1 (25.00%). Vals called it weaker on Terminal-Bench 4.0 (39.39%) and Tax Agent Bench (38.23%).

Simon Willison drew his usual pelican on a bicycle at all six reasoning levels. The high setting failed to return a complete SVG, and the drawing did not get steadily better as reasoning went up. He judged the max-effort pelicans comparable to those from Muse Wonder, which costs 40 times as much per token. One drawing proves little. Still, it is a reason to test lower settings before you assume max is best.

Meta’s launch post also quotes five early partners: Box, Cline, Oracle, Cloudflare and You.com. Cloudflare’s chief technology officer says Ember ran up to 30% faster at the tail than the comparable model it tested, and You.com reports a 95th-percentile response time of 2.2 seconds on its coding tasks. Meta chose these quotes, and none comes with a published method.

How did Muse Ember do in our agency test?

On four everyday agency jobs, Muse Ember at high effort matched GPT-6.1 Terra’s accuracy for about half Terra’s cost, and it was the only model that never misread a figure on a badly compressed screenshot. It did not beat GPT-6 Luna, which costs the same per token: on these short jobs Luna at medium effort was at least as accurate on search terms and landing pages and cheaper still, though it misread more figures.

On 12 November 2026 we ran the same test we used for GPT-6.1 Terra, with the same inputs and answer keys, through OpenRouter. Ember ran on Meta’s own endpoint at four effort settings; Luna and Terra ran on OpenAI’s. Each set-up did every job twice: 728 requests, which cost $0.70 in all.

If the table extends beyond the screen, scroll sideways to view all columns.

Model and effort	Search terms labelled right	Kept searches marked negative	Landing-page checks right	Hard screenshot values misread	Cost of the test	Seconds per batch of search terms
Muse Ember, minimal	Search terms labelled right89.5%	Kept searches marked negative5	Landing-page checks right99.4%	Hard screenshot values misread0	Cost of the test$0.03	Seconds per batch of search terms2.9
Muse Ember, medium	Search terms labelled right93.5%	Kept searches marked negative1	Landing-page checks right98.8%	Hard screenshot values misread0	Cost of the test$0.08	Seconds per batch of search terms7.2
Muse Ember, high	Search terms labelled right95.3%	Kept searches marked negative1	Landing-page checks right98.8%	Hard screenshot values misread0	Cost of the test$0.10	Seconds per batch of search terms9.5
Muse Ember, max	Search terms labelled right95.0%	Kept searches marked negative2	Landing-page checks right93.8%	Hard screenshot values misread0	Cost of the test$0.15	Seconds per batch of search terms17.7
GPT-6 Luna, medium	Search terms labelled right96.8%	Kept searches marked negative5	Landing-page checks right100%	Hard screenshot values misread6	Cost of the test$0.03	Seconds per batch of search terms11.2
GPT-6 Luna, max	Search terms labelled right96.3%	Kept searches marked negative6	Landing-page checks right99.4%	Hard screenshot values misread6	Cost of the test$0.11	Seconds per batch of search terms73.3
GPT-6.1 Terra, medium	Search terms labelled right95.3%	Kept searches marked negative2	Landing-page checks right100%	Hard screenshot values misread2	Cost of the test$0.20	Seconds per batch of search terms8.5

Ember’s effort setting made the clearest difference on search terms. At minimal effort it labelled 89.5% right and gave the same answer in both runs for only 81% of terms. At high it reached 95.3%, level with Terra, and agreed with itself on 92%. Max added no accuracy, cost 54% more than high and took nearly twice as long per batch.

The mistakes that cost money are searches the client wants to keep but the model marks as negative keywords, which would stop the ads showing to real buyers. Ember at medium or high made that mistake once in 400 answers. Luna made it five or six times at either setting, and Terra twice.

Screenshots showed Ember’s caution. Every set-up read all 116 values on our eight standard dashboard screenshots correctly and made up nothing for the values we had covered. The hard set was decided by one report saved as a heavily compressed JPEG. On it, Ember flagged the same four visible figures as unreadable in every run and never guessed. Luna misread six values, reading a £41.39 cost per conversion as £41.70 or £61.79, and Terra read 544 clicks as 344 in both runs. For dashboard reporting, a model that says it cannot read a figure is safer than one that guesses.

Ember’s one stumble came at max effort. On one landing-page screenshot it thought for 793 tokens and then returned an empty answer, so all eight checks on that page counted as wrong. Without that reply it would have scored 158 of 160, like medium and high.

The cost result runs against the headline. Ember thought far longer than Luna at the same setting name: at medium effort it used 110,540 reasoning tokens across the test, against 23,691 for Luna at medium. At the same token price, that made Ember at medium 2.8 times as expensive as Luna at medium on our jobs. Effort names such as “medium” do not buy the same amount of thinking from different makers, so compare models on cost per finished job.

This is a small test on synthetic material, so treat gaps of a point or two as noise. Every job was a single short request, so it says nothing about long agent loops, tool calls or coding, where Artificial Analysis and Meta put Ember’s biggest advantage. Claude, made by Anthropic, built the test and drafted this post, so we did not test Claude Haiku 5.5.

Can Muse Ember work with images, audio and video?

Meta says yes to all three. Its model page and docs list text, image, video, audio and PDF input, and the launch post says agents can work from screenshots, scanned documents, voice memos, meeting recordings and video without a separate model. OpenRouter lists the same inputs.

The independent testers list less. Artificial Analysis records text and image input. Vals lists text, image and file input and marks video input as not supported. Those pages describe what each tester measured or could use, so they do not prove the other inputs fail. But we found no published outside results for Ember on audio or video.

Image work has the best evidence. Meta’s PerceptionBench score of 80.6% with tools is its own test, but its photo example below put 10,000 photos through Ember. Treat images and PDFs as ready to trial. Test audio and video on your own files before you build anything on them.

How should you set up Muse Ember?

Use Ember as a worker under a planning model, give each worker a small, fully specified job, and turn the effort setting down where the job is simple. Meta’s developer blog shows the pattern with two worked examples.

In the first, Muse Wonder planned and supervised a port of 16 language lexers in Pygments, the Python syntax highlighter, from Python to Rust. Sixteen Ember agents each took one lexer in its own copy of the repository. Meta reports that every port matched Python on 383 test inputs and on 62 held-out files, and that the Rust versions ran 3.0 to 7.3 times faster than the Python originals. Ember handled 158 million input tokens to Wonder’s 4.4 million, and the run cost about $3.96: $2.12 for Wonder and $1.84 for Ember.

In the second, Wonder chose eight albums from a sample of photos and wrote one script, and 16 Ember agents captioned, renamed and sorted 10,000 photos. Meta says the albums matched the dataset’s own labels for about 86% of photos. The post gives two totals for the run, $1.59 and $1.33, and in both Ember’s share is about 50 cents. Meta estimates the same photo requests would have cost about $20 at Wonder’s prices.

Five settings decide whether Ember stays cheap in practice:

1. Set the effort level yourself. Ember defaults to max. In our test, high effort labelled search terms as well as max for two-thirds of the cost and half the wait, while minimal made five times as many costly mistakes. Meta’s own photo pipeline ran Ember at minimal effort on 512-pixel images, so try the lowest setting that passes your checks.
2. Keep each worker’s context under 200,000 input tokens. Past that line the whole request costs three times as much, and a stored conversation counts towards it.
3. Hand over the full brief. Meta’s worker prompts name the files the worker may change, the checks it must run and the result to return, and send any wider problem back to the planner.
4. Ask for structured output on routing and extraction, so every answer arrives in a schema your code can check.
5. Reuse your OpenAI code. The Meta Model API accepts OpenAI-compatible requests, so trying Ember usually means a new base URL, a new key and the model ID muse-ember-1.0.

Can UK teams use Muse Ember, and what happens to their data?

Yes, UK developers can use it. Meta’s Geographic Use Policy, last updated on 9 November 2026, lists the territories where the Meta Model API is barred, and the UK is not among them. Its only model-specific restrictions cover Muse Wonder and Muse Wonder Full in South Korea, and the same update lists Ember as a Standard model.

The data terms depend on how you reach the model:

• Meta Model API, pay as you go: Ember runs in Standard mode only, where Meta says your prompts, files and the model’s answers are not used to train its models. Unlike Muse Spark, Ember has no cheaper Contributor variant.
• Muse Code subscription: the policy says subscribers use Contributor mode by default, whichever model they pick, which lets Meta train on their content. Switch it off in /settings on each computer you use before you point Ember at client code. Our Muse Code guide walks through the settings.
• OpenRouter: Ember is served there by Meta at the same token prices; the Google Vertex route on OpenRouter now works only with your own Google Cloud key. OpenRouter says the model is not currently available through its EU in-region routing, so teams that rely on that option for data location cannot use it for Ember yet.
• Google Cloud and Oracle Cloud: Meta says Ember, Spark and Wonder are now available on both, which suits teams that already buy AI through one of those clouds.

Whichever route you choose, check the provider’s processor terms against your UK GDPR obligations before you send personal data through an agent built on Ember.

Should you switch to Muse Ember?

If you run Claude Haiku 5.5 or GPT-6 Luna at high effort inside agent loops, test Ember now. It costs the same per token, a third of Haiku’s cost per task at max effort on Artificial Analysis’s runs, and it starts answering in seconds where Haiku at max takes minutes. Set it to high effort: in our test that matched max on accuracy, cost two-thirds as much and never produced the empty reply we saw at max.

For short, single-step jobs such as sorting search terms, the case is weaker. In our test GPT-6 Luna at medium effort was at least as accurate on search terms and landing pages and cost less than a third as much as Ember at high, though it marked more searches the client wanted to keep as negatives and misread six figures on a blurred report. If misread numbers are your biggest risk, Ember’s habit of flagging what it cannot read is worth paying for. If cost per request is, start with Luna.

Keep a bigger model for jobs where one hard answer matters more than the bill. Haiku 5.5 beats Ember on FrontierCode 2.0, and Meta itself sends hard single-shot coding to Muse Spark or Muse Wonder. Run GPT-6.1 Terra in the same trial if you use OpenAI: at medium effort it scores two points higher on Artificial Analysis’s index for three cents more per task, though in our test it cost twice as much as Ember at high.

Ember suits Meta’s own planner-and-worker set-up best. If you already use Muse Wonder or Spark in Muse Code, Ember comes with every plan, and Meta’s examples show it cutting the cost of large parallel jobs to a few dollars. If you need a model you can run on your own servers, look at Muse Glimmer or Muse Spark Open instead, because Ember’s weights are private.

Before you switch anything in production, run a few hundred of your own real tasks through Ember at two effort levels and compare accuracy, cost per finished job and time with your current model. Our test took under an hour and cost less than a dollar. If you want help setting up that comparison, talk to AIWIZ.

Sources

• Introducing our smallest, most efficient model: Muse Ember, Meta AI Research
• Muse Ember evaluation report, Meta AI Research
• Start building with Muse Ember, Meta Model API blog
• Muse Ember model page and Muse Ember docs, Meta Model API
• Pricing and rate limits, Meta Model API
• Model Designation and Geographic Use Policy, Meta Model API
• Muse Ember (max) and LLM leaderboard, Artificial Analysis
• AutomationBench-AA, Artificial Analysis
• Muse Ember benchmarks, Vals AI
• Meta: Muse Ember, OpenRouter
• Muse Ember, Simon Willison
• Meta releases Muse Ember, a new lightweight model for agent execution, TechCrunch
• Pricing, Anthropic
• GPT-6 Luna model page, OpenAI