November 9, 2026, evening
Meta launches Muse Ember, which it calls its smallest model yet, with one line that shows up in every headline: Claude Haiku 5.5's level, at a third of the cost.
I opened the three price lists. Muse Ember: 10 cents per million input tokens, 50 for output. Claude Haiku 5.5: 10 and 50. GPT-6 Luna: 10 and 50.
Three models, the exact same price. So where does the "third" come from?
Muse Ember is Meta's small, cheap model, released on November 9, 2026. Meta built it for the day-to-day work of AI agents: sorting, pulling data out of documents, using tools and browsing websites, at volume. I went through the official pages, the independent measurements from Artificial Analysis and the first developer tests, to answer one question: does this launch change anything for a business that runs an agent, or is it just another model with a nice chart?
We didn't run a test of our own. The model is a day old, and every number here comes from open sources, linked at the end.
In short
- Muse Ember costs exactly the same per token as Claude Haiku 5.5 and GPT-6 Luna: $0.10 per million input tokens and $0.50 for output.
- The "third" is cost per task: on Artificial Analysis's independent index it matches Haiku 5.5's score at 14 cents a task vs. 41, and finishes in 3.2 minutes vs. 14.8.
- The weak spots: general knowledge (ask it general facts from memory, with no document in front of it, and it declines to answer 64% of the questions), harder code, and, in one independent test, a habit of saying "done" when the action failed.
- My take: a significant launch if you run a lot of routine work through an agent. Not a main model, and no reason to replace a model that already works well for you.
Prefer to watch? The review in a 9-minute video
My take, at a glance
My take: a real launch, with a misleading headline
Muse Ember didn't lower the price. It lowered the bill. The price floor of 10 and 50 cents was set by GPT-6 Luna in September and Claude Haiku 5.5 in October. Meta arrived third at the same price. What it brought along: a model that reaches the same result with less text and in less time.
Think of three plumbers who charge the same hourly rate. One of them finishes the same repair in a third of the time, so his bill is three times smaller. That's Ember. And the analogy has a less pleasant sequel: in one test, that same plumber also said "fixed it" while the tap was still dripping, more often than his rival. We'll get to that.
So, significant or hype? Significant, in a narrow area. If your agent does hundreds of small actions a day, reading emails, tagging inquiries, filling in fields, this launch is worth a look. If you're after a model that thinks for you, writes strategy or fixes a tricky bug, this launch doesn't concern you.
What Muse Ember is, in a minute
Sources: Meta's model page and pricing page; speed per Artificial Analysis, November 10, 2026. We explain tokens in What is a token.
It's available through Meta's direct connection to its models, the API (the way an agent works with a model), on every Muse Code plan (Meta's coding tool), on Google Cloud and on OpenRouter, a service that connects you to many models through one account. Meta doesn't mention it in the Meta AI app or in WhatsApp, so if you don't run an agent or write code, you won't come across it directly. It's a closed model: Meta doesn't release the model file, and you can't run it on your own server. Israel is not on the API's list of restricted countries.
What about languages other than English? Meta doesn't publish which languages it tested, and every test I found was in English. Before you hand it emails or reviews in another language, test it on your own real examples.
Meta presents it as the other half of Muse Wonder, its large model: Wonder plans, Ember handles the volume. According to Meta, it's also its first model trained with the help of Wonder Full.
One detail that matters for the bill: Ember has six reasoning levels, meaning how long it "thinks" before answering, from minimal to max, and the default is max. You can't switch reasoning off. If you connect it and change nothing, you pay for the longest thinking on every request, even when the task is tagging an email. Meta itself recommends the lowest level that gives a reasonable result.
Muse Ember vs. Haiku 5.5 and GPT-6 Luna: same price list, different bill
Artificial Analysis is an independent company that runs the same set of tests on every new model and publishes an overall score out of 100, plus cost and time per task. For scale: Muse Wonder, Meta's large model, scores 44 there. It's the most thorough independent measurement of Ember published so far, and most of the numbers here come from it.
A "task" on this index is a heavy test task, with lots of reading and thinking. Tagging an email or pulling the total from an invoice costs a fraction of that. What matters in the chart is the ratio between the models, not the cents.
Cost per task on the Artificial Analysis index, and overall score
Source: Artificial Analysis, index version 5, November 10, 2026. DeepSeek is calculated at its peak-hours price. Version 5 scores are lower than the previous version's and can't be compared with numbers in earlier articles.
Against Haiku 5.5: the same score, 36. Before rounding, Haiku is slightly higher: 36.2 vs. 35.8. The difference is in how they get there: to reach that score, Haiku produced about 2.6 times as much text. And when the price per token is identical, whoever writes less costs less.
Against GPT-6 Luna: almost the same cost per task, 14 cents vs. 15, but Ember scores 36 and Luna 28. Here the advantage is quality, not money.
Inside Meta: Muse Spark 1.3, its cheapest option until this launch, scores 37 at $2.39 per task. Ember scores one point lower and costs 17 times less.
There's one more reason the bill is smaller, and it matters especially for agents. An agent sends the model the same text over and over: its instructions, the conversation history, the document it's working on. Text the model has already read (cached input) is billed at a discount, and with Ember that's $0.004 per million tokens, vs. $0.01 for Haiku and Luna. According to Artificial Analysis, on its agent tests that's most of the cost. DeepSeek is even cheaper on this line, but it writes far more text per task, so it ends up costing more.
Average time per task, in minutes
Source: Artificial Analysis, November 10, 2026. All models at maximum reasoning.
Speed is its clearest advantage, and Meta barely talked about it. Ember starts answering in under 5 seconds, even at maximum reasoning. Haiku 5.5 at the same setting thinks for three minutes before the first word. At low reasoning levels Haiku and Luna also answer within seconds, but then their scores drop a lot. Ember is the only one of the three that gives both the high score and a fast answer, and that's what matters for an agent replying to a customer in chat.
Where Meta's numbers look too good
Four things that aren't in the launch headline:
1. The comparison was made only at maximum reasoning. At that setting Haiku writes the most, so it's the most expensive. One level lower, Haiku scores 35 and costs 25 cents per task. The real gap, if you tune your model properly, is about 1.8x. Still an edge for Ember, but not a third.
2. Some of the wins are on tests Meta ran itself. Three of the tests in the launch table are Meta's own internal ones. And on the computer-use test, the score it shows, 69.5%, includes points for partial success. Count only tasks completed all the way, and the score is 33.8%. Haiku drops the same way, to 27.8%, so Ember still leads, but it's far from the impression that seven out of ten tasks succeeded.
3. On one test, Haiku clearly wins. FrontierCode checks whether code the model wrote would be accepted in a real project. Haiku: 34.9%. Ember: 27.2%. Meta shows this row in its own table, to its credit.
4. The price step at 200K tokens. Meta sells a one-million-token window, but once the input passes 200K tokens (a few hundred pages of English text), the price triples, and on the whole request, not just the part above the line. The input also includes the agent's instructions and the conversation history. In Ember's favor: Haiku has a steeper step, 5x, starting as early as 100K. Above 272K, GPT-6 Luna is cheaper than Ember.
And one more thing, which sounds like an advantage: Ember hardly ever makes up answers. Its rate of made-up answers is 17.4%, about half of Haiku's. But that comes at a price. Artificial Analysis asked the models factual questions from memory, with no document in front of them, and Ember declined to answer 64% of them. Overall it answered 21% of the questions correctly, vs. 35% for Haiku. For an agent working with your documents and data, that's good behavior: "I don't know" beats a made-up answer. If you want to ask it questions from memory, it's a weak model. We explain what a hallucination is and why it happens in What is an AI hallucination.
Real tests: what people found when they ran it
There isn't much material yet. These tests were published in the first hours after launch, mostly by individual developers or companies that sell tools for agents. Anthropic's Claude Fable 5.2 also came out the same day, and most of the attention went there.
Four tests, and the results
Green: Ember is better value. Orange: depends on what matters to you. Sources: kritinv07, browser_use, navaneeth_pk and LatentOperatorX on X, November 9–10, 2026.
The test that matters most for a business: customer service. Developer Kritin Vongthongsri ran 400 simulated service chats on each model, for a store and an airline, each task eight times. On first-try success it was a tie: 83.2% for Ember, 82.8% for Haiku. The difference showed up in the failures. In 24 of Ember's 50 failures, it told the customer the action had succeeded. In one example it processed an exchange instead of a return, and wrote to the customer that the exchange had gone through. With Haiku, this happened in 31% of failures. It's one test, on 50 failures, and the tester himself writes that on overall reliability neither model clearly leads. More tests are needed to know whether it's a pattern, but it's exactly the kind of mistake a business can't afford.
And the bill: Ember's chats ran longer, 22.5 messages on average vs. 16.6. When you work out what an answer you can trust costs, Ember came to 22 cents and Haiku to 18. Same price list, and Ember is 20% more expensive.
Kritin Vongthongsri on X: Muse Ember vs. Claude Haiku 5.5 on 800 service chats, November 10, 2026
Browsing websites: this is where it shines. Browser Use, which makes a tool that lets an agent browse in a web browser, ran it on tasks of hundreds of steps. It got the same score as GPT-6 Luna at medium reasoning, at about 3 cents a task. This is a company that sells a tool for agents, so it has an interest in cheap models doing well.
Browser Use on X: Muse Ember on a test of long browser tasks, November 9, 2026
Against DeepSeek: faster, less pretty. A developer built ten internal apps with each of three models, and had reviewers rate them without knowing which model built what. The price came out the same. Ember finished three times faster than Luna, at similar or better quality, according to him. DeepSeek V4.1 Flash scored 8.1 on design, vs. 6.9 for Ember. The developer appears to be affiliated with the platform the apps were built on.
Navaneeth PK on X: Muse Ember vs. GPT-6 Luna and DeepSeek V4.1 Flash, November 10, 2026
And two warnings from the field. A Japanese developer found that on his test Ember got a perfect score, but wrote about three times as much text as Muse Spark and took longer. And a developer on Hacker News who worked with the trial version before launch wrote that it starts strong, and by the twentieth message in a conversation it's no longer usable. That's one account of an early version, and according to the Latent Space newsletter, problems raised during the trial period were fixed before launch. It does fit the way Meta presents it: a model for repetitive work at volume, not for long conversations.
Muse Ember pros and cons
What you get
- Haiku 5.5's score at a third to half the cost per task
- Speed: answers start within seconds, even at maximum reasoning
- Text it has already read, like the agent's standing instructions, costs 60% less than with Haiku and Luna
- Reads images, video, audio and PDF
- Would rather say "I don't know" than make things up
- In the API, your content isn't used for training
What you give up
- Weak general knowledge: not a model to ask questions from memory
- Complex code: Haiku 5.5 is better
- In one test it reported "done" on actions that failed, more often than Haiku
- Above 200K input tokens, the price triples
- Closed model: runs only in the cloud of Meta and its partners
- No discount for tasks that don't need an instant answer, unlike Haiku and Luna
A quick word on privacy, because it's confusing. In the API, according to Meta's terms, your content isn't used to train its models. On a Muse Code plan, the default is the opposite: your content is used for training, with whichever model you choose, until you turn it off in the settings. Turning it off applies only from that moment on, and anything already sent stays available for training. And if you reach Ember through Google Cloud, Meta says it doesn't receive your content at all. We covered privacy in Meta's consumer product in Muse privacy.
Use cases: what I'd hand Muse Ember in a business agent
- Sorting emails and website inquiries by urgency
- Pulling the amount, date and supplier from invoices
- Tagging Google reviews: positive, negative, needs a reply
- Image descriptions and tags for products on your site
- Collecting prices and information from websites
- Customer service that takes actions: cancellations, refunds, order changes
- Filling in forms in your business systems
- Small fixes to your website's code
- Factual questions from memory, with no document it can read
- Marketing copy and decisions
- A tricky bug or a big change to your site
- A long conversation of dozens of messages
- A document of hundreds of pages at once
My rule: Ember gets tasks that have a beginning, an end and a way to check they were done. "Pull the total from this invoice": yes. "Handle this customer": no, unless a stronger model or the system itself checks that the refund really went through. The false "done" finding is the reason: when you deal with customers, a mistake reported as a success costs more than anything you save.
If your agent runs on Haiku 5.5 or Luna and works well, there's no reason to rush to switch. The difference is cents per task, and switching models means retesting everything the agent does. Ember is worth testing in two cases: when your agent does thousands of actions a day and the bill is already noticeable, or when speed matters, for example live replies to customers. If someone manages your agent for you, ask whether they can run Ember on one of your tasks alongside your current model and compare the result and the bill. And if your agent runs on Google's model, do the math now: Gemini 3.8 Flash costs $0.75 and $3.75, and its price doubles on January 1, 2027.
What I'm doing with it
I usually build agents on Hermes Agent, an open-source framework from Nous Research, and pick the models that fit each business: from OpenAI, Anthropic, Google, Meta and others. Platin Digital is an OpenAI Select Partner, and that doesn't change the choice. Most agents use more than one model: a strong model that makes the decisions, and a cheap model that does the day-to-day work. Ember is a serious candidate for the second role, and I'll wait for more independent measurements, mainly on reliability when using tools and on long conversations, before putting it in front of a business's customers.
Within seven weeks, OpenAI, Anthropic and Meta landed on exactly the same price in the cheap tier, and Google went the other way. We break down what an agent costs beyond the model price in How much an AI agent costs a business, and the full model ranking is in the Sigma AI Ranking.
Frequently asked questions
What is Muse Ember?
Muse Ember is Meta's smallest and cheapest model, released on November 9, 2026. It was built for the day-to-day work of AI agents at high volume: using tools, pulling data out of documents, sorting, browsing websites and small coding tasks. It reads text, images, video, audio and PDF, answers in text, and is available in Meta's API as muse-ember-1.0, in Muse Code, on Google Cloud and on OpenRouter.
How much does Muse Ember cost?
$0.10 per million input tokens, $0.50 for output, and $0.004 for text the model has already read. That's the same price as Claude Haiku 5.5 and GPT-6 Luna, except for already-read text, which is 60% cheaper with Ember. When the input passes 200K tokens, the whole request is billed at three times the rate: $0.30 and $1.50. Meta doesn't publish a free tier in the API, and there's no discount for tasks that don't need an instant answer.
Is Muse Ember better than Claude Haiku 5.5?
On the Artificial Analysis overall score they're tied, at 36. Before rounding, Haiku is slightly higher. Ember is three times cheaper per task at maximum reasoning and 4.6 times faster. Haiku 5.5 is better at factual questions from memory and at complex code, and in one customer-service test it less often reported success on something that had actually failed. For routine work at volume, Ember is better value. For work that needs knowledge and accuracy with customers, Haiku.
Is Muse Ember available in Israel?
Israel isn't on the list of restricted countries for Meta's API, and the model is also available through Google Cloud and OpenRouter. We didn't test sign-up and payment from Israel ourselves. It isn't part of the Meta AI app or WhatsApp, so it's a model for people who run an agent or write code. We covered the availability of Meta's consumer agent in Muse in Israel.
Is my data used to train Meta's models?
Not in the API: according to Meta's terms of service, content sent to Muse Ember isn't used to train its models. On a Muse Code plan, the default is that your content is used for training, and you can turn it off in the settings. Meta doesn't publish a fixed retention period for data. If you use it through Google Cloud, Meta says it doesn't receive the content at all.
What's the difference between Muse Ember, Muse Spark and Muse Wonder?
These are Meta's three tiers. Muse Wonder is the large model, $4 and $20 per million tokens, for complex tasks and planning. Muse Spark 1.3 is the middle tier, $1.25 and $4.25. Muse Ember is the small one, $0.10 and $0.50, for doing work at volume. On the Artificial Analysis index, Ember scores 36, one point below Spark 1.3, at a cost per task 17 times lower.
Sources
- https://research.meta.ai/blog/muse-ember
- https://dev.meta.ai/resources/blog/build-with-muse-ember
- https://dev.meta.ai/docs/muse-ember
- https://dev.meta.ai/docs/pricing-rate-limits
- https://dev.meta.ai/docs/reasoning
- https://dev.meta.ai/docs/choose-a-model
- https://research.meta.ai/static/muse-ember-evaluation-report
- https://dev.meta.ai/legal/terms-of-service
- https://dev.meta.ai/legal/geographic-use-policy
- https://cloud.google.com/products/model-garden/meta
- https://artificialanalysis.ai/models/muse-ember
- https://artificialanalysis.ai/articles/meta-s-muse-ember-is-the-cheapest-model-scoring-36-on-the-intelligence-index
- https://artificialanalysis.ai/leaderboards/models
- https://platform.claude.com/docs/en/about-claude/pricing
- https://developers.openai.com/api/docs/pricing
- https://ai.google.dev/gemini-api/docs/pricing
- https://api-docs.deepseek.com/quick_start/pricing
- https://venturebeat.com/technology/metas-proprietary-muse-ember-nearly-matches-claude-haiku-5-5-at-1-3-the-estimated-per-task-cost
- https://thenextweb.com/news/meta-muse-ember-small-ai-model-haiku-luna
- https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/
- https://siliconangle.com/2026/11/09/meta-introduces-muse-ember-to-power-cost-efficient-ai-agents/
- https://news.ycombinator.com/item?id=50429024
- https://news.ycombinator.com/item?id=50434838
- https://x.com/kritinv07/status/2119981282281345112
- https://x.com/browser_use/status/2119922023975829744
- https://x.com/navaneeth_pk/status/2119980781506568501
- https://x.com/LatentOperatorX/status/2120014259186966703
- https://x.com/gosrum/status/2119956559031927079
- https://openrouter.ai/meta/muse-ember
- https://latent.space/p/ainews-fable-5-2-retakes-world-1-llm-spot-as-ant-ipo-roadshow-begins
