 |
| AI-generated illustration of a small glowing ember flame hovering above a compact AI chip, with glowing circuit traces spreading across a dark desk. |
On Monday 9 November 2026, Meta's Superintelligence Labs released Muse Ember, the smallest and most efficient model in its Muse family — and the cheapest. It is not a frontier model. It will not write you a novel, solve a physics problem, or top the benchmark charts. It is built for a different job entirely: the repetitive work AI agents do all day, thousands of times over. Tool calls, coding subtasks, routing requests, pulling data out of long documents. Meta says Ember matches Anthropic's Claude Haiku 5.5 on one major benchmark index at about a third of the cost per task, and TechCrunch covered the launch the same day. This guide explains what Ember is, what the cost claim actually means, where the small model falls short, and who should care about it.
Two cautions before the detail, because a cheap model is easy to misunderstand. First, the benchmark figures in this article are Meta's own, drawn from version 5 of the Artificial Analysis Intelligence Index, which was published the same day. Meta says Ember delivers more intelligence per dollar than Claude Haiku 5.5 or OpenAI's GPT-6 Luna at every reasoning setting, and that no model on the efficiency frontier scores higher for less per task. Those are Meta's readings of a third-party index, and the cost-per-task figures are estimates at Meta's list prices, not measured bills from real users. Second, Ember is a specialist with a known weak spot: its knowledge score is the lowest among the models Meta compares it with, and it declines nearly two-thirds of trivia-style questions rather than answering them. A model that is cheap and fast is not the same as a model that knows a lot. Keep those two limits in mind and Ember reads as what it is: a very good small model for a specific kind of work, not a small model that does everything.
As an Amazon Associate, we earn from qualifying purchases.
Prices and savings can change. Check the retailer for the current price.
What Ember is for: the work agents repeat all day
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 list is the key to understanding the model. A modern AI agent does not make one big decision; it makes hundreds of small ones. It calls a tool, reads the result, decides what to do next, calls another tool, summarises a document, routes a request to the right handler, and repeats the loop. Most of those steps do not need a frontier model. They need a model that is fast, cheap, and reliable enough to run thousands of times a day without the bill exploding.
That is the job Ember is built for. It is the volume worker in Meta's lineup, and Meta says so directly: "Muse Wonder plans, Muse Ember handles the volume." The pattern is a pairing. A large model like Muse Wonder sets the plan and the rules; a small model like Ember applies them across thousands of cases. TechCrunch gives the example of a customer-support agent: a heavier model establishes the rules for handling a particular type of problem, then a lighter-weight model applies those rules across tens of thousands of cases. The same pattern works for coding agents, where a big model designs a change and a small model runs the tests, makes the small edits, and extracts the results.
This is also why Ember's price matters more than its raw intelligence. When a model runs hundreds of times inside a single agent job, the cost per call is the whole business model. A model that is slightly less capable but ten times cheaper can be the difference between an agent product that makes money and one that does not.
The price, and the cost-per-task claim
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 list prices on the Meta Model API, and they make Ember significantly cheaper than frontier competitors. For comparison, Claude Haiku 5.5 charges the same standard rate per token, but Meta says Ember uses fewer tokens per job, which is where the savings come from.
The headline claim is about cost per task, not cost per token. On the Artificial Analysis Intelligence Index, Meta says Ember scores 36 at its maximum reasoning setting — the same rounded score as Claude Haiku 5.5 at its best — at an estimated $0.14 per task, against $0.41 for Haiku. That is roughly a third of the cost for the same benchmark score. Meta also says Ember beats GPT-6 Luna at every reasoning setting, and that no model on the full Artificial Analysis leaderboard scores higher for less per task. The gap widens further against Meta's own larger models: Muse Spark 1.3 scores 37 on the same index, one point higher than Ember, but costs $2.39 per task — about seventeen times as much.
Read the fine print, though. The $0.14 and $0.41 figures are Artificial Analysis estimates based on its prompts and measured token use at list prices; your production bill will vary with the shape of your workload. Cached input is nearly free at $0.004 per million tokens, which rewards workloads that repeat the same context, but punishes nothing in particular — it simply means the savings are biggest when the same long prompt is reused many times. And the comparison is against Haiku's best score at its highest reasoning setting, which is also its most expensive setting. The claim is real, but it is a benchmark claim, not a guarantee about your bill.
The specifications: a million tokens in, half a million out
Ember's technical specifications are generous for a small model. It has a 1 million-token context window, enough to take in a full codebase or a long document set in one call, and it can return up to 512,000 tokens in a single response. It accepts text, images, audio and video as input, which Meta says lets agents read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video, without switching to a separate model.
Ember is available now in Muse Code, Meta's coding agent, and through the Meta Model API as muse-ember-1.0, with support for the Responses API. It is also listed on OpenRouter, and Meta says the Muse models are now available on Google Cloud and Oracle Cloud as well. Meta has not disclosed a parameter count, and it has not said in which countries the Meta Model API is available. The weights stay private, as they did with the earlier Muse releases — with one exception, the open-weights Muse Glimmer. If you were hoping to download Ember and run it yourself, that is not on offer.
Meta also notes that Ember is the first model it trained using Muse Wonder Full, its largest model. That detail matters because it explains the strategy: the biggest model teaches the smallest one, and the smallest one does the volume work. It is the same pattern the industry is converging on — frontier models as teachers, small models as workers — and Ember is Meta's clearest expression of it so far.
Where Ember falls short
No honest guide to Ember can skip its weaknesses, and Meta's own published figures make them easy to find. The most important is knowledge. On the Artificial Analysis AA-Omniscience test, Ember scores 21.1 percent, the lowest among the similarly intelligent models Meta compares it with, against 34.7 percent for Haiku 5.5. Artificial Analysis says knowledge is Ember's main weakness, and one analysis found it declines 64 percent of trivia-style questions rather than answering them. A model that often says "I don't know" is safer than one that guesses, but it is also less useful when the job needs general knowledge rather than a document in front of it.
Ember also trails Haiku 5.5 on some coding tests. On FrontierCode 2.0, Ember scores 27.2 percent against Haiku's 34.9 percent. It leads on other agent benchmarks — AutomationBench-AA and the Coding Agent Index-AA — so the picture is mixed rather than uniformly behind, but the gap on FrontierCode is a reminder that "matches Haiku" depends on which test you look at. The Intelligence Index is a composite; it hides as much as it reveals.
There are practical limits too. Ember has no consumer rollout: it lives in Muse Code and the Meta Model API, and Meta has not said when, or whether, it will reach the consumer Muse app. There is no parameter count, no open weights, and no published list of countries where the API is available. And the safety picture, while reassuring on paper — Meta says its pre-release assessments placed Ember below the high-risk thresholds in its Superintelligence Scaling Framework for chemical and biological, cybersecurity, and loss-of-control risks — is Meta's own assessment, not an outside audit.
The Muse family, four models in
Ember is the fourth model released under the Muse brand, and the family has grown fast. It began with Muse Spark in April 2026, a mid-weight model. Spark's open-source twin, Glimmer, followed in August. The Wonder frontier model arrived in October. Ember, in November, completes the set: a small, cheap model for volume work, sitting underneath the frontier model that plans the work and the mid-weight model that handles coding and agentic workflows.
The consumer side of Muse has moved just as fast. The flagship Muse Agent launched in September 2026 to what TechCrunch describes as unexpected success, with new features like Spaces and video chat rolling out earlier this month. That success is the context for Ember: Meta needs cheap, fast models to serve millions of agent interactions, and the Muse Agent's growth is what makes a model like Ember commercially necessary rather than merely interesting.
It is also worth noting what day Ember launched on. Monday 9 November was a big day for AI models: Anthropic released Claude Fable 5.2 and Mythos 5.2 the same day, taking the top spot on the Artificial Analysis index. The two launches are a study in contrast — Anthropic pushing the frontier outward, Meta pushing the cost of volume work down — and together they show where the industry is heading: not one model to rule them all, but a fleet of models, each priced for the job it does.
Who should try Ember
- Developers building AI agents. If your agent makes hundreds of tool calls, routing decisions or extraction passes per job, Ember is built for exactly that workload. Test it on your most repetitive subtask first.
- Teams with high-volume, cost-sensitive workloads. Long-document parsing, bulk extraction, and classification are the jobs where Ember's price and 1M-token context pay off fastest.
- Anyone already using Muse Code. Ember is available there now, so you can try it without changing your setup or signing up for a new service.
- Test it against Haiku 5.5 on your own tasks before switching. The benchmark claim is strong, but your workload is the test that matters. Run both models on the same jobs and compare cost and quality.
- Do not use it for knowledge-heavy work. Ember's knowledge score is its weakest point. If the job needs general knowledge rather than a document in front of it, use a larger model.
- Watch the cached-input pricing. At $0.004 per million cached tokens, workloads that reuse the same long context get the biggest savings. Design your prompts to take advantage of it.
- Do not wait for a download. The weights are private. If you need a model you can run yourself, look at the open-weights Glimmer instead.
Bottom line
Muse Ember is not the most powerful model Meta has ever released, and it is not trying to be. It is the cheapest, and that is the point. At $0.10 per million input tokens, with a 1 million-token context window and a benchmark score that matches Claude Haiku 5.5 at about a third of the cost per task, it is built for the unglamorous work that makes AI agents affordable: tool calls, routing, extraction, and long documents, repeated thousands of times a day. Its weaknesses are real — thin general knowledge, a mixed coding record, closed weights, and no consumer app yet — but they are the weaknesses of a specialist, not a failure. The right question is not whether Ember is the best model, but whether it is the best model for the repetitive job you need done a thousand times. For a growing number of developers, the answer is now yes.
Reference sources
- Meta, "Introducing Our Smallest, Most Efficient Model: Muse Ember", 9 November 2026: research.meta.ai
- TechCrunch, "Meta releases Muse Ember, a new lightweight model for agent execution", 9 November 2026: techcrunch.com
- Unite.AI, "Meta Releases Muse Ember for Lower-Cost Agent and Coding Work", 9 November 2026: unite.ai
- AlphaSignal, "Meta Releases Muse Ember to Cut AI Agent Costs by 3x", 9 November 2026: alphasignal.ai
- CellCog, "Muse Ember: Meta's Smallest Model, Price and Benchmarks", 9 November 2026: cellcog.ai
Leave a public comment