Meta has a new answer to the most expensive question in AI: what does it cost to run an agent all day? On November 9, 2026, Meta’s Superintelligence Labs launched Muse Ember, its smallest and most efficient model yet — a lightweight model built for the repetitive work AI agents do constantly, at a price that makes running agents at scale look affordable for the first time.
The headline number is striking. Ember matches Claude Haiku 5.5, Anthropic’s leading small model, on the Artificial Analysis Intelligence Index — both score 36 — but Meta says it does each benchmark task for about $0.14, against $0.41 for Haiku. That is roughly a third of the cost for the same headline score. It is the clearest sign yet that the AI race is splitting in two: frontier models chasing the hardest problems, and small, cheap models doing the bulk of the everyday work.
What Is Muse Ember?
Muse Ember is the fourth model in Meta’s Muse lineup, following the mid-weight Spark (released in April 2026), Spark’s open-source twin Glimmer (August) and the frontier Wonder model (October). Its model ID is muse-ember-1.0, and it is available now on the Meta Model API and in Muse Code, Meta’s coding agent. It is also listed on OpenRouter and available on Google Cloud and Oracle Cloud.
Meta describes Ember as designed for “complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.” That is a deliberately narrow brief. Ember is not meant to be the model that plans your hardest problem. It is meant to be the model that carries out the hundreds of small steps that follow — reading files, running tests, calling tools in sequence, pulling records from one app and updating another.
It is also the first model Meta trained using Muse Wonder Full, its largest model — a teacher-student pipeline in which the frontier model helps build the smaller one. Meta says it has even more efficient models in training.
The model is multimodal by design. It natively understands text, images, audio and video, so an agent built on Ember can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video without handing off to a separate model. It ships with a 1-million-token context window and can produce up to 512,000 tokens of output in a single call — enough to take in a full codebase or document set at once. It supports six reasoning levels, from minimal to max, so developers can trade intelligence against cost and speed per task. It also supports the Responses API.
The Price: Cheap per Token, Cheaper per Task
Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. Those are the same per-token rates as Claude Haiku 5.5 and OpenAI’s GPT-6 Luna, which is exactly why Meta’s pitch is about cost per task, not cost per token: Ember uses fewer tokens to reach the same result, so the bill per completed job comes out lower.
The comparison Meta published is blunt. On the Artificial Analysis Intelligence Index at maximum reasoning, Ember costs about $0.14 per task, Haiku 5.5 costs $0.41 and GPT-6 Luna costs $0.15. Against Meta’s own mid-weight model the gap is wider still: Muse Spark 1.3 scores 37 on the same index — one point higher than Ember — but costs $2.39 per task, roughly 17 times as much, at $1.25 per million input tokens and $4.25 per million output tokens.
There is one pricing catch worth knowing before you build on it. The headline rates only hold for requests up to 200,000 input tokens. Above that threshold, all three rates triple — to $0.30 per million input tokens, $0.012 cached and $1.50 per million output tokens — applied to the whole request. Long-context work is still cheap, but it is not three-times-cheaper the way the headline suggests.
The Benchmarks: Where Ember Wins — and Where It Doesn’t
Meta published a detailed benchmark table comparing Ember at maximum reasoning with Haiku 5.5 and GPT-6 Luna, using Artificial Analysis evaluations from November 9, 2026. On the headline index, all three models at their highest reasoning setting score 36, 36 and 28 respectively. But the individual tests tell a more interesting story, and it is not a clean sweep.
Ember’s strongest results are in agent and coding work. On AutomationBench-AA, which runs end-to-end business workflows across real applications, Ember scored 37.8 percent against 13.9 percent for Haiku and 20.2 percent for Luna — nearly triple Haiku’s score. On Terminal-Bench 4.0, an agentic terminal coding test, it scored 40.9 percent against Haiku’s 39.2 percent and Luna’s 16.4 percent. On the Coding Agent Index, which measures coding inside real agent harnesses, it scored 46 against 37 and 31. And on Agents’ Last Exam, a set of end-to-end agent tasks, it reached 61.6 percent against 57.0 and 57.1.
It also leads on computer use and perception. On OSWorld 2.1, which tests operating desktop applications from screenshots, Ember scored 69.5 percent on partial credit against Haiku’s 68.9 percent and Luna’s 50.5 percent. On PerceptionBench, a fine-grained visual perception test with tools, it scored 80.6 percent against 78.6 and 62.8. On Meta’s internal Personal Superintelligence Index, a composite of personal and knowledge work, it scored 46.8 against 36.1 and 37.0. On long-document work, the GDP.pdf-AA test, it tied Luna at 22.8 percent, ahead of Haiku’s 20.8.
The counter-example matters too. On FrontierCode 2.0-AA, which judges whether generated code is good enough to merge into a real project, Haiku leads clearly at 34.9 percent, with Ember at 27.2 and Luna at 25.2. The lesson is the one Meta itself states in its positioning: for the hardest problems, pair Ember with Muse Wonder — Wonder plans, Ember handles the volume. Ember is built for repeatable work at scale, not for the toughest single task.
One caveat belongs with every number in this section. These are Meta’s published results, drawn from Artificial Analysis evaluations and Meta’s own internal tests. They have not been independently replicated, and rival models were run in different harnesses on some rows. Treat them as Meta’s claims about its own model, not as settled independent findings. That framing matters even more in a week when Anthropic launched Claude Fable 5.2 with its own record benchmark claims — every vendor’s launch-day chart flatters its own model.
Built for Agents: Tool Calls, Coding Subtasks and Long Documents
Agents spend most of their tokens on many small, structured steps, and Ember is built for exactly that pattern. Meta’s pitch is that an agent can carry a business workflow end to end on Ember — pulling records, calling tools in sequence and updating systems across apps — without the cost of each step adding up to a bill that kills the use case.
For coding, the target is the subtask an agent sends off hundreds of times in a session: reading files, running tests, making targeted fixes. The Terminal-Bench and Coding Agent Index results above are Meta’s evidence that a small model can do that work well, and the pricing is what makes it practical — at $0.14 per benchmark task, an agent that fires off hundreds of calls a day stops being a luxury.
The multimodal side extends the same idea beyond code. Because Ember reads images, audio and video natively, one model can work through a scanned contract, a recorded meeting or a screenshot of a broken interface. Meta points to long real-world documents as a particular strength: with a million-token context window, the model can take in the whole document set in one call rather than chunking it and losing context between pieces.
Early Testers: Box, Cline, Oracle, Cloudflare and You.com
Meta named five early testers that put Ember to work before launch: Box, Cline, Oracle, Cloudflare and You.com. Their use cases sketch the target customer clearly — enterprise content, coding tools, cloud platforms and search — rather than consumer chatbots.
Box’s verdict is the most detailed. Yashodha Bhavnani, VP of AI Products at Box, said Ember “can score a contract against a company’s negotiation playbook clause by clause and land on the right overall risk rating, or screen candidates against a detailed brief and flag the borderline ones who still qualify as alternates,” calling it “a fast, efficient foundation for agents doing real work on enterprise content.” You.com reported that Ember finished every coding task it set, with a 95th-percentile response time of 2.2 seconds — the kind of latency number that matters when an agent is waiting on dozens of model calls in a row.
Safety and Availability
Before release, Meta says it evaluated Ember under its Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks, and found it remained below the framework’s high-risk thresholds. Though it is a small model, Meta says it received the same core safety training and testing as its larger siblings.
The company also evaluated how Ember behaves when acting on a user’s behalf — whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages, documents or tool outputs that try to redirect it, and seeks human input before high-impact actions. That last point matters for agents specifically: a model that runs tools all day needs to know when to stop and ask.
Ember is available today on the Meta Model API and in Muse Code as muse-ember-1.0, and on Google Cloud and Oracle Cloud. Meta has not released the model weights — this is an API-only model, not an open download — which puts it in a different category from the open-source Glimmer twin of Spark.
What It Means for Developers and Businesses
For developers, Ember changes the maths of building agents. The cost of running an agent has been the quiet barrier to shipping one: a capable model firing hundreds of calls per workflow adds up fast, and most teams responded by limiting what their agents were allowed to do. A model that matches Haiku’s headline score at a third of the cost per task — and that beats it outright on business-workflow and coding-agent tests — removes much of that constraint, at least for the repetitive steps.
The practical pattern Meta recommends is a team of models: Wonder plans the hard parts, Spark handles serious coding and agentic workflows, and Ember carries the volume. That division of labour is becoming the industry’s default shape. The same week Ember launched, OpenAI’s Businesses in ChatGPT turned the chatbot into a place where businesses deploy agents to talk to customers, and Anthropic’s new office suite pushed AI deeper into everyday office work. Agents are no longer a demo category; they are the product. Ember is priced to be the engine inside them.
For small businesses, the appeal is direct. The tools that let a small team act bigger — AI tools for small businesses that save time and money — get cheaper when the model underneath them costs cents per task. A shop that could not justify an agent that reads every incoming email, extracts the order, checks stock and drafts a reply might now be able to. The 200,000-token price step is the thing to watch: long-document workflows that cross it will cost three times the headline rate, so teams should measure their real prompt sizes before budgeting.
The honest limits are worth stating plainly. Ember is not the model for the hardest single problem — Wonder and Spark exist for that, and Haiku still beats Ember on mergeable-code quality. Its benchmarks are Meta’s own, not independently verified. And it is API-only, so teams that need to run a model on their own infrastructure will look elsewhere. But for the work agents actually repeat all day, the combination of Haiku-level scores and a $0.14 task cost is the most practical AI launch of the week.
Bottom Line
Muse Ember is Meta’s bet that the future of AI is not one giant model doing everything, but a team of models with the cheapest one doing most of the work. At $0.10 per million input tokens and $0.50 per million output tokens, with a million-token context window and Haiku-matching scores at about a third of the cost per task, it makes agent workloads affordable at a scale where they previously were not. It is available now on the Meta Model API and in Muse Code. Start with Ember for the repetitive steps, reach for Wonder when the problem is genuinely hard — and keep an eye on the 200,000-token threshold where the price triples.
Frequently Asked Questions
What is Muse Ember?
Muse Ember is Meta’s smallest and most efficient AI model, launched November 9, 2026. It is built for the repetitive work AI agents do all day: tool calls, coding subtasks, routing, extraction and long-document parsing.
How much does Muse Ember cost?
It costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, for requests up to 200,000 input tokens. Above that threshold, all three rates triple.
How does Muse Ember compare with Claude Haiku 5.5?
Both score 36 on the Artificial Analysis Intelligence Index at maximum reasoning, but Meta says Ember costs about $0.14 per benchmark task against $0.41 for Haiku. Ember leads on business workflows and coding-agent tests; Haiku leads on FrontierCode mergeable-code quality.
Where can I use Muse Ember?
It is available on the Meta Model API and in Muse Code as muse-ember-1.0, and on Google Cloud and Oracle Cloud. It is also listed on OpenRouter. Meta has not released the model weights, so it is API-only.
What is Muse Ember best for?
Repetitive agent work at scale: tool calls, coding subtasks, routing requests, extracting data from long documents, and operating computers from screenshots. For the hardest problems, Meta recommends pairing it with Muse Wonder.
Is Muse Ember safe to use in agents?
Meta says it evaluated Ember under its Superintelligence Scaling Framework and found it below high-risk thresholds, with the same core safety training as its larger models, including tests for resisting hidden instructions in web pages and documents.
