What Is Muse Ember? Meta’s Smallest Model, Its Price, and How to Use It

Muse Ember is Meta’s smallest and most efficient AI model: a system built for the repetitive work AI agents do all day, rather than the hardest single question you can ask it. Announced on Monday 9 November 2026 by Meta Superintelligence Labs, it is the fourth model in the Muse family — and the cheapest per task at its capability level, matching Claude Haiku 5.5 on the Artificial Analysis index at about a third of the cost per task.

The simplest way to understand Ember is through the job it was made for. Modern AI agents do not answer one question and stop. They run loops — read a file, call a tool, check the result, call another tool — hundreds of times in a session. Each step is small, and each step costs tokens. A model slightly less brilliant per step but dramatically cheaper per step wins that kind of work, and that is the bet Meta is making.

What is Muse Ember?

Ember is a proprietary reasoning model, identified as muse-ember-1.0, available through Meta’s Model API and inside Muse Code, Meta’s coding agent. It works through a problem with extended chain-of-thought before answering. Meta calls it a model for “complex reasoning execution” — execution is the operative word. Ember is not the model you ask to invent a new strategy; it is the model that carries out the hundreds of small steps a strategy requires.

It accepts text, images, audio and video as input, and produces text. That breadth matters, because the material agents work with is rarely clean text: screenshots, charts, scanned documents, voice memos and video — all taken in directly, without a separate model to transcribe them first. Its context window holds one million tokens, with up to 512,000 output tokens, so one call can take in a full codebase or a long document set. It also supports Meta’s Responses API, built for multi-step agent workflows that keep state between calls.

What Ember is not: it is not open. Meta has not disclosed its parameter count, and its weights are not publicly available — unlike Glimmer, released as an open-source twin of Spark. Ember can only be used through Meta’s API, Muse Code, or cloud partners. There is no model file to download, and no way to run it on your hardware.

The Muse family: Spark, Glimmer, Wonder, Ember

Ember only makes sense inside the family Meta has built around it. The line began with Spark, a mid-weight model released in April 2026 for coding and agentic workflows. In August came Glimmer, an open-source twin of Spark, and Muse Code, the terminal agent that runs on these models. In September, Meta released the Muse Agent, its flagship consumer product. In October came Wonder, the frontier model for the hardest problems — and Wonder Full, the most powerful version, for Muse Code subscribers.

Ember, released in November, completes the range. Meta’s summary is blunt: Wonder is for the hardest problems, Spark is for coding and agentic workflows, and Ember is for repeatable coding and agentic tasks at scale. A customer-support agent, for example, might have Wonder establish the rules for handling an issue, then hand them to Ember to apply across tens of thousands of cases. Meta’s shorthand: “Muse Wonder plans, Muse Ember handles the volume.”

Ember is also the first model Meta says was trained using Muse Wonder Full — a sign that Meta now uses its flagship as a teacher to build cheaper, specialised models, a pipeline that could produce further small models. Meta says it has “even more efficient models in training.” For a closer look at another small model built for long, unsupervised work, see our guide to Inkling 2 Small, Thinking Machines Lab’s 594B-parameter model, released the same week with a very different philosophy.

What “agent execution” means in practice

Meta lists five kinds of agent execution: tool calls, coding subtasks, routing, extraction, and long document parsing. In each, the task is well-defined, the volume is high, and the cost of each call matters more than the difficulty of any one call. Tool calls are the clearest example: an agent that books travel or updates a CRM spends most of its tokens on structured steps — reading a tool’s output, deciding which tool to call next, and formatting the arguments. None of those steps needs frontier reasoning, but all need reliability, because one malformed call breaks the workflow. Routing sends a request to the right specialist or system. Extraction pulls structured facts out of unstructured material. Long document parsing is the heavy end: reading a hundred-page document and answering questions about it, or scoring it against a checklist.

The early testers Meta published show what this looks like in practice. Box used Ember to score a contract against a company’s negotiation playbook clause by clause, and to screen job candidates against a detailed brief. Cloudflare used it for investigations spanning several code repositories, reporting it ran up to 30% faster on its slowest responses. Cline expects to use Ember as a subagent — the small model doing routine steps while a larger model supervises. Oracle reported encouraging results on complex engineering tasks. You.com measured a p95 latency of 2.2 seconds and says Ember used 36–48% fewer output tokens than the leading open-weights models it compared against. These are Meta-selected testimonials, not independent audits — but they point at the same conclusion: the model’s value shows up in throughput and latency on routine work, not in a single dazzling answer.

Pricing and the cost-per-task maths

Ember’s price list is short: $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, all pay-as-you-go. The cached-input price is the number to notice: agents re-read the same context — the same codebase, playbook, or conversation — on every step, so cached input often dominates the bill. Re-reading a million cached tokens costs less than half a cent. Artificial Analysis calculates a blended rate of about $0.0728 per million tokens across a typical mix of cached, fresh-input, and output tokens.

The more useful figure is cost per completed task, because that is what a business actually pays for. Artificial Analysis measures the weighted average cost of running one task from its Intelligence Index: about $0.14 per task for Ember, about $0.41 for Claude Haiku 5.5, and about $0.15 for GPT-6 Luna. Ember’s sibling, Muse Spark 1.3, costs about $2.39 per task — roughly seventeen times Ember’s figure — for one extra point on the index. These are estimates at list pricing, and they will move with caching and task mix, but the ordering is the point: Ember is in a different price band from rivals at the same capability level.

To make that concrete: a support agent handling 10,000 cases a day, at one index task per case, would cost about $1,400 a day on Ember at list prices, about $4,100 on Haiku 5.5, and about $23,900 on Muse Spark 1.3. The arithmetic is illustrative — real bills depend on caching and output volume. One warning: Ember is somewhat verbose, generating 115 million output tokens across the index against a 90-million median for its price tier, and output tokens are the expensive kind. The per-task figure already accounts for that, but anyone estimating from input prices alone will undercount. Speed reinforces the throughput story: Artificial Analysis records 3.1 minutes per index task and 4.22 seconds to first answer token, both among the quickest in its price tier.

The benchmark evidence — and its limits

Meta published a full comparison table. On the Artificial Analysis Intelligence Index, Ember scores 36, matching Haiku 5.5’s best setting and ahead of Luna’s 28. The index is independently measured, and Artificial Analysis’ own model page confirms the 36, placing Ember “well above average among other models in a similar price tier” — the median in that tier is 24.

The agent-specific benchmarks are where Ember is strongest. On AutomationBench, an end-to-end test of business workflows, it scores 37.8 against 13.9 for Haiku and 20.2 for Luna — nearly triple Haiku’s score. On Agents’ Last Exam it leads 61.6 to 57.0 and 57.1, and on coding it leads Terminal-Bench 4.0 (40.9 to 39.2 and 16.4) and the Coding Agent Index (46 to 37 and 31). On computer use it scores 69.5 against 68.9 and 50.5, and on visual perception with tools it leads 80.6 to 78.6 and 62.8. Meta also reports an internal composite, the Personal Superintelligence Index, at 46.8 against 36.1 and 37.0 — but that is Meta’s own test, with no outside confirmation.

The honest reading has to include the two places Ember does not lead. On FrontierCode 2.0, which tests real engineering tasks judged on whether the work could actually be merged, Haiku leads 34.9 to Ember’s 27.2, with Luna at 25.2. On GDP.pdf, professional work over long documents, Ember ties Luna at 22.8, ahead of Haiku’s 20.8. Neither gap is fatal, but they mark the boundary of what it is for. Three caveats apply to the whole table. First, the chart is Meta’s presentation of Artificial Analysis data at list pricing, with each rival at its best reasoning setting; the index scores are independently measured, but the framing is Meta’s. Second, benchmark tasks are not your tasks. Third, the most striking results are from new tests, and outside researchers have not yet run their own head-to-head comparisons.

How to access Muse Ember

Ember is available now, with no waitlist mentioned. The main route is the Meta Model API, where the model ID is muse-ember-1.0. It is also inside Muse Code, and supports the Responses API for multi-step workflows. Meta also says Ember, Spark and Wonder are now available on Google Cloud and Oracle Cloud, so teams on those clouds can call it without a separate Meta account setup.

For developers, the practical starting point is Muse Code if you already use it, since Ember slots into the same agent loop as Spark. In a harness, the natural pattern is the one Meta describes: a larger model plans, Ember executes — route routine subtasks such as file reads, test runs, targeted fixes and extraction to muse-ember-1.0, and reserve the expensive model for planning and final review. There is no free tier mentioned in the announcement, and no regional availability list: Meta does not say in which countries the Model API is available for Ember, so check availability in your region before building on it. For a closer look at another agent-style launch from the same period, see our guide to Perplexity Teammates, the AI agents for work.

How Muse Ember compares

Against Claude Haiku 5.5, the comparison Meta wants you to make, Ember is a near-tie on capability and a clear win on price: both score 36 on the Artificial Analysis index at their best settings. Ember leads on business workflows and agentic coding; Haiku leads on the hardest engineering tasks. Per completed task, Ember costs about a third of Haiku — routine volume favours Ember, hard one-off engineering favours Haiku. Against GPT-6 Luna, Ember leads on both fronts: Luna scores 28 against Ember’s 36, and Meta says Ember beats Luna at every reasoning setting. Per task, Luna costs about $0.15 against Ember’s $0.14, so the price gap is small but the capability gap is not. Against Muse Spark 1.3, Ember is the value choice: Spark scores 37, one point higher, but costs about $2.39 per task against $0.14. Spark is for workflows where the extra point is worth the premium; Ember is for the same workflows repeated at scale. Against open-weights models in its price range, such as DeepSeek V4.1 Flash and GLM 5.3 Flash, Ember is cost-competitive as a hosted, managed model — no self-hosting, no infrastructure — which trades control for convenience.

Who should use it — and who should wait

Ember fits teams running agents at volume: support agents handling thousands of similar cases, coding agents that read files and run tests hundreds of times a session, and document pipelines extracting structured data — any workflow where the same context is re-read repeatedly and cached pricing pays off. It also fits builders who want a cheap subagent inside a larger harness, with a frontier model supervising. Others should look elsewhere. If your work is mostly hard, one-off reasoning — novel engineering, open-ended research, strategy — Wonder or a frontier rival is the better tool, and Ember’s FrontierCode gap to Haiku warns against that use. If you need to self-host, fine-tune freely, or run a model offline, Ember cannot: it is proprietary and API-only, and Glimmer is the open option. If you need audio output or image generation, Ember produces text only. And if your tasks are short and chat-like rather than long agent loops, the cost advantage shrinks, because there is less repeated context to cache.

Safety and alignment: what Meta disclosed

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 below the framework’s high-risk thresholds, with the same core safety training as its larger models. The agent-specific checks are the interesting part: Meta says it tested how Ember behaves when acting on a user’s behalf — whether it respects intent, handles sensitive information, resists instructions hidden in web pages, documents, or tool outputs, and seeks human input before high-impact actions. That last category matters most for Ember’s work, since an agent that reads the web and runs tools all day is constantly exposed to text written to manipulate it. Meta’s disclosure says the model was tested on this; it is not a guarantee, and no numbers were published.

What to watch next

The first thing to watch is whether the efficiency claim survives contact with real users: the test that matters is whether developers find Ember reliable across weeks of repetitive work. Independent head-to-head testing is the next milestone — none has been published yet. The second is what Meta builds with the Wonder Full training pipeline: Ember is the first model trained using Wonder Full, and Meta says more efficient models are in training. The third is the response from rivals: Anthropic and OpenAI both have lightweight models in the same segment, and a price cut from either would narrow Ember’s advantage directly. The fourth is availability — which regions and clouds get Ember, whether a free tier appears, and whether it reaches the other agent platforms, such as Cline and Cursor, that decide which models developers actually encounter.

For now, Muse Ember is best understood as a price breakthrough with a capability story attached. It does not move the frontier of what AI can do — it moves the frontier of what AI can do cheaply, at volume, all day. For the agents that will do most of the world’s routine AI work, that may matter more.

Related news coverage: our news report on the Muse Ember launch