Muse Ember: Meta’s Cheap Model That Beats Haiku on Cost
Meta has launched Muse Ember, a small, efficient model aimed at the least glamorous and most numerous part of the artificial intelligence workload. It targets the repetitive steps that make up an agent's day. Fetch the customer record, check the document, run the test, update the ticket, call the next tool. None of that needs a frontier model, and Meta is pricing Muse Ember as if none of it should pay for one.
The numbers behind the launch are more interesting than the usual benchmark table. Artificial Analysis, which runs independent evaluations, found that Muse Ember at maximum effort is the cheapest model scoring 36 or higher on its Intelligence Index, at 0.14 dollars per task. Claude Haiku 5.5 scores the same 36 at 0.41 dollars per task, nearly three times the cost, on identical input and output token prices. That difference comes from efficiency rather than sticker price, and it is the whole story of this release.
What Muse Ember is
Muse Ember comes from Meta Superintelligence Labs and is positioned as the efficient tier of the Muse family. It handles text, images, audio and video as input, and carries a context window of up to 1 million tokens. It is available through the Meta Model API under the identifier muse-ember-1.0, in Muse Code, and through Google Cloud and Oracle Cloud. Meta's research blog positions it as the company's smallest and most efficient model to date.
The pricing is aggressive on paper: 0.10 dollars per million input tokens, 0.004 dollars per million cached input tokens, and 0.50 dollars per million output tokens, with a higher tier for prompts beyond 200,000 tokens. Those input and output rates are the same as Claude Haiku 5.5 and GPT-6 Luna. The advantage shows up in the details. Cached input, which dominates the cost of agentic evaluations, is 60 percent cheaper than the competitors'. The model also uses roughly a third as many output tokens as Claude Haiku 5.5 to finish the same tasks.
Meta's positioning is explicit about the division of labour. Muse Ember is for high-volume, lower-complexity work: chained tool calls, document extraction, small coding tasks, reading files, running tests, applying specific changes. For complex projects, Meta suggests pairing it with its larger Muse Wonder model, with Wonder planning and reasoning while Ember executes. That is the standard frontier-plus-worker architecture, and Muse Ember is Meta's entry in the worker category.
What the independent numbers say
Artificial Analysis ran Muse Ember through its Intelligence Index and agentic suites, and the results are notably mixed in an instructive way.
At the top level, Muse Ember (max) scores 36, matching Claude Haiku 5.5 (max) while costing about a third as much per task. Muse Ember (xhigh) scores 34 at 0.12 dollars. It is roughly four times cheaper than DeepSeek V4.1 Flash 1017, which scores 33 at 0.60 dollars, and GLM-5.3-Flash at 28 and 0.58 dollars. Against its own family, it sits 8 points behind Muse Wonder, which scores 44 at 6.95 dollars per task, about one fiftieth of Ember's cost.
The agentic results are where the model earns its positioning. On AutomationBench-AA, which measures automated workflows, Muse Ember scores 68.1 percent, level with Muse Wonder and GPT-6 Astra. It beats Claude Haiku 5.5 on GDP.pdf, a benchmark of long professional documents, at 22.8 percent to 20.8 percent, though it trails on AA-Briefcase Elo, 1521 to 1574.
Coding is genuinely mixed. Muse Ember beats Haiku on Terminal-Bench 4.0, 40.9 percent to 32.8 percent, but loses on FrontierCode 2.0, 27.2 percent to 34.9 percent. In its own Muse Code agent, however, it scores 46 on the Coding Agent Index, ahead of Haiku running in Claude Code at 37 and GPT-6 Luna in Codex at 31. The pattern suggests the model performs best inside its own toolchain, which is convenient for Meta and worth knowing for anyone evaluating vendor claims.
The knowledge trade, stated honestly
The most interesting number in the independent evaluation is also the least flattering. On AA-Omniscience, which measures factual accuracy and hallucination, Muse Ember scores 21.1 percent accuracy, the lowest of the similarly intelligent models compared, against 34.7 percent for Claude Haiku 5.5 and 50.5 percent for Gemini 3.8 Flash.
The reason is a design choice rather than a failure. Muse Ember declines to answer 64 percent of knowledge questions, and that refusal behaviour keeps its hallucination rate at 17.4 percent, half of Claude Haiku 5.5's. The model would rather say nothing than invent something.
For the intended workload, that is arguably correct. An agent spending its day fetching records and running tests gains little from encyclopedic knowledge and suffers a great deal from confident fabrication in tool chains, where one invented fact propagates through every downstream step. A model that demurs is safer in a pipeline than a model that improvises. For anyone planning to use Muse Ember as a knowledge source, though, the evaluation says plainly that this is not that product. Pairing it with a search tool or a larger model for factual questions is not optional, it is the design.
Why the cost-per-task framing matters
The conventional way to compare models is price per million tokens, and by that measure Muse Ember is unremarkable: its rates match its competitors. The Artificial Analysis evaluation compares cost per completed task, and that is where the gap appears.
The mechanics are straightforward. In agentic workloads, the dominant cost is cached input, because every step of a task re-sends the accumulated context. Muse Ember's cached input price is 60 percent cheaper than the models it competes with. It also finishes tasks with about a third of the output tokens, which suggests terser reasoning and less verbosity in tool calls. Multiply those two effects across a thousand-task workload and the difference between 0.14 dollars and 0.41 dollars per task becomes the difference between 140 dollars and 410 dollars, every day, at scale.
That framing also clarifies who the model is for. Organisations running a handful of complex tasks per day will not notice the difference. Organisations running thousands of agent steps per day, which is what the automation trend is producing, will notice it immediately. Muse Ember is priced for the second group. The fact that a hyperscaler-scale competitor's list price now has a credible cost-per-task challenger is a sign of where competition in the model market is heading, away from headline benchmarks and toward the invoice.
Inside the efficient tier's economics
The efficient-model tier has its own logic, and it is worth spelling out because it differs from frontier competition in almost every way.
Frontier competition is decided by capability: whichever model tops the hardest benchmarks, whatever the cost. Efficient competition is decided by economics: whichever model completes a standard task for the least money at an acceptable quality floor. The benchmarks shift accordingly, from reasoning exams to agentic suites, from peak scores to score per dollar, and from single-shot answers to long tool chains where cached context dominates the bill.
That shift changes what model builders optimise. Muse Ember's cited advantages, cheaper cached input and fewer output tokens, are not capabilities in the usual sense. They are behaviours: terser reasoning, less verbose tool output, better use of caching. Those behaviours can be trained. A model that learns to solve a task in fewer turns is worth more to an agent operator than one that scores a point higher on an exam while tripling the token bill.
It also changes what buyers should measure. A procurement evaluation of an efficient-tier model should run the buyer's own tasks at the buyer's own volume and count the invoice. A published benchmark cannot capture how chatty a model is with a particular toolset. The Artificial Analysis cost-per-task numbers are a useful starting point, and the purchase decision should still end with the buyer's own meter.
There is a downside to watch as this tier grows. Efficiency pressure can reward models that skip work rather than do it poorly. The knowledge-refusal behaviour that makes Muse Ember safe in pipelines is the same behaviour that makes it useless as a source of facts. Tier separation is healthy as long as buyers understand which tier they are buying.
The safety posture
Meta reports that Muse Ember was tested under its Superintelligence Scaling Framework despite its size, covering cybersecurity, chemical and biological misuse, and loss-of-control risks, and that the model stayed clear of high-risk thresholds in the categories tested.
The filing also addresses agentic behaviour specifically. The assessments covered how the model acts on behalf of users: protecting sensitive information, resisting instructions embedded in external documents and websites, and handling misleading content in the material it processes. Meta says the model is designed to seek human approval before taking actions with significant consequences. That posture aligns with the guardrails the industry has been adding to agent platforms, from agent workspaces with supervised delegation to the governed agent features shipping in mainstream developer tools.
The computer-use and multimodal angle
Muse Ember's input list, text, images, audio and video, is not decoration. It maps directly onto the computer-use workload that agentic products are converging on.
A screenshot is how an agent sees a screen, and the model's visual input supports reading interfaces, extracting values and deciding what to click, which is what computer-use agents actually do. Audio and video input serve the extraction workload: processing a recording into notes, pulling figures from a webinar, transcribing and summarising without routing through a second model. The announcement's example workload, an assistant that fetches customer data, reviews paperwork, categorises inquiries and updates systems, is exactly the multi-format mix those inputs serve.
The context window is the other half of that story. A 1 million token window means the model can hold a large document collection or a long session of tool results without external retrieval on every step. The 512,000-token range is what Meta highlights for large document collections and codebases. Whether the efficient tier degrades gracefully at those depths is a question deployment will answer, but the headroom exists.
That last design choice pairs naturally with the model's knowledge behaviour. A model that declines to answer uncertain questions and asks for approval before consequential actions is a model built to be delegated to carefully. Whether those behaviours hold up under adversarial pressure is the kind of question only deployment will answer, but the posture is coherent, and it matches what enterprises have been asking for from worker-tier models.
Where it leaves the Muse family, and the market
Muse Ember completes a Meta lineup that has been building all year. Muse Wonder put Meta back at the frontier, and Muse Spark demonstrated that Meta's open weights could run on ordinary hardware. With Ember, the family now covers frontier reasoning, open-weight deployment and efficient execution, which is the full spectrum that Anthropic and OpenAI each cover with more models.
The competitive picture is sharp. Haiku 5.5 and GPT-6 Luna hold the same token prices and lose on cost per task. DeepSeek and the Chinese efficient models hold lower scores. The efficient tier is becoming a distinct market with its own economics, judged on dollars per completed task rather than on peak capability, and Meta has entered it with the current best claim on that metric.
Distribution is the second front. Muse Ember is available on Google Cloud and Oracle Cloud as well as Meta's own API. That means Meta is selling its efficient tier through competitors' clouds, the same playbook OpenAI executed with its AWS Marketplace listing and the same logic the hyperscalers adopted when they started hosting rival models. Distribution through incumbents has become the default strategy for every major lab, and the efficient tier is where that strategy pays, because efficient models are the ones enterprises buy by the millions of tasks.
The remaining question is durability. Efficient-tier leadership changes hands quickly, because the recipe, aggressive pricing plus cost-focused training, is replicable, and DeepSeek's pricing already undercuts several western models. Muse Ember holds the cost-per-task claim today. Whether it holds it in a quarter is the actual competitive question, and the answer will arrive in someone else's evaluation.
The caveats should be kept alongside the claim. The 36 Intelligence Index score is real and independently measured, and so is the weak knowledge performance. A buyer whose workload involves answering factual questions should look elsewhere or add retrieval. A buyer whose workload is automation should model the cost difference on their own task volume, because at agent scale, the efficiency advantage is not a rounding error, it is the product.
Who should look elsewhere
The efficiency pitch has clear limits, and three workloads should not be routed to Muse Ember without a second thought.
Factual question answering belongs elsewhere, because the AA-Omniscience result shows the model declining most knowledge questions and its remaining answers are right barely a fifth of the time. Frontier reasoning belongs elsewhere, because 8 points below Muse Wonder is a real gap on hard problems, and Wonder exists precisely for them. And high-stakes judgement calls, legal interpretations, medical decisions, financial commitments, belong with a model whose knowledge base is deeper and whose licence to be wrong is smaller.
The efficient tier is not a lesser version of the frontier tier. It is a different product with a different contract: less knowledge, less invention, less cost, more willingness to ask for help. Buyers who match the workload to the contract get the savings. Buyers who do not get exactly what the evaluation predicts.
Questions readers are asking
How much does Muse Ember cost?
0.10 dollars per million input tokens, 0.004 dollars per million cached input tokens, and 0.50 dollars per million output tokens, with a higher tier beyond 200,000-token prompts at 0.30, 0.012 and 1.50 dollars respectively. Artificial Analysis measured 0.14 dollars per task at max effort on its agentic evaluations.
How does Muse Ember compare with Claude Haiku 5.5?
On independent evaluation, it matches Haiku 5.5's Intelligence Index score of 36 at roughly a third of the cost per task. It beats Haiku on Terminal-Bench 4.0 and AutomationBench-AA, trails on FrontierCode 2.0 and AA-Briefcase, and has substantially weaker factual knowledge with half the hallucination rate.
What is Muse Ember for?
High-volume, lower-complexity agentic work: chained tool calls, document extraction, small coding tasks and automated workflows. Meta suggests pairing it with Muse Wonder for complex projects, with Wonder planning and Ember executing.
Is Muse Ember multimodal?
Yes. It accepts text, images, audio and video as input, which covers the screenshot-reading workload of computer-use agents and the extraction workload for recordings and scanned documents. Output is text.
How much context can Muse Ember handle?
The context window reaches 1 million tokens, and Meta highlights the 512,000-token range for large document collections and codebases. Prompts beyond 200,000 tokens are billed at a higher rate.
Does the weak knowledge score matter?
It depends on the workload. For automation, the low hallucination rate is a feature, because fabricated facts poison downstream steps. For factual question answering, the 21.1 percent accuracy means the model will often decline rather than answer, and a retrieval tool or larger model should handle knowledge instead.
The bottom line
Muse Ember is not trying to win benchmarks that make headlines, and that is what makes it interesting. It is built for the part of the AI economy that is actually growing, the millions of small steps agents take every day. The independent numbers say it does that work at a third of the cost of the cheapest comparable competitor.
The trade is knowledge, and it is stated in the numbers. The model declines most of what it does not know and rarely invents, which is the right behaviour for a worker in a pipeline and the wrong behaviour for an oracle. Meta's efficient tier now has a credible claim on the best cost-per-task in its class. The efficient-model competition, unlike the frontier race, is one where a price list and an independent evaluation can settle the argument.