Meta Muse Ember: Pricing, Benchmarks and How to Use It

Meta announced Muse Ember on November 9, 2026, describing it as its smallest and most efficient model. It is available in Muse Code and the Meta Model API as muse-ember-1.0, with a 1M-token context window, up to 512K tokens of output, and Responses API support. For prompts up to 200K tokens it costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, a price point aimed at high-volume, repeatable agent work.

The short version: Muse Ember is not built to maximize raw intelligence, it is built to minimize the cost of each task. It scores 36 on the Artificial Analysis (AA) Intelligence Index versus 37 for Muse Spark 1.3, yet costs about $0.14 per index task against $2.39 for Spark 1.3, roughly 17 times less. It suits subagents, repeated tool calls, routing, extraction and long-document parsing. Meta's blog does not state the parameter count or whether the weights are open, so it cannot be treated as a model you run locally.

Meta's announcement positions Muse Ember for complex reasoning execution and repeated agent work: tool calls, coding subtasks, routing, extraction and long-document parsing. In Meta's lineup, Muse Wonder handles the hardest problems, Muse Spark covers coding and agentic workflows, and Muse Ember covers repeatable tasks at scale.

It is the first model trained using Muse Wonder Full. The blog gives no other training details, no parameter count and no statement about open weights. Meta says it natively handles text, images, audio and video, including operating a computer from screenshots and reading scanned documents, charts, audio and video. AA's model page lists text and image input with text output, so check your provider's exact modality support before building on audio or video.

- The first model trained with Muse Wonder Full, scoring 36 on the AA Intelligence Index at a small-model price
- A 1M-token context and up to 512K tokens of output at $0.10 per million input tokens
- Meta claims more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, on the efficiency frontier of the AA leaderboard
- Available through Muse Code and the Meta Model API, plus Google Cloud Model Garden and Oracle Cloud
- Under the Meta Superintelligence Scaling Framework, Meta reports results below the high-risk thresholds for chemical and biological, cybersecurity and loss-of-control categories

All of this rests on Meta's own announcement and AA's published figures, not on independent reproduction.

The Meta Model API charges by prompt length, per million tokens.

Past 200K tokens, input, cached input and output all triple. You can feed a full 1M-token document set, but at a higher unit price. Cached input costs 4% of regular input (about 96% off), so agents that reuse a long shared prompt such as policies, a codebase or tool definitions pay even less. AA puts the blended rate at about $0.073 per million tokens using a 7:2:1 mix of cached input, input and output.

Muse Spark 1.3 on the Standard tier costs $1.25 input and $4.25 output, so Ember is 12.5 times cheaper on input and about 8.5 times on output. On AA's cost per index task, Ember at max reasoning is about $0.14 and Spark 1.3 at max is $2.39. The Spark 1.3 Contributor tier, which requires sharing your data, matches Ember's $0.10 input price.

The table below shows the blog's results at max reasoning against Claude Haiku 5.5 and GPT-6 Luna. Figures are as published by Meta unless noted, and metrics marked Meta-internal were designed and run by Meta.

Three things stand out. Ember leads the agent benchmarks and Terminal-Bench 4.0, but Terminal-Bench uses each model's best published harness (mini-swe-agent for Ember, Claude Code for Haiku, Codex for Luna), so conditions differ. On FrontierCode 2.0-AA, Claude Haiku 5.5 clearly leads at 34.9% against Ember's 27.2%, a 7.7-point gap on code mergeability. And on GDP.pdf-AA, Ember only ties Luna at 22.8%.

OSWorld 2.1, PerceptionBench, Agents' Last Exam and the Personal Superintelligence Index are Meta-internal, so outsiders cannot check whether they favor Meta's models. Lean on the AA-derived metrics and test on your own workloads.

AA's model page lists Muse Ember (max) at 36 on the Intelligence Index, ranked second of 33 models. It outputs 446 tokens per second, takes 4.37 seconds to the first answer token and 3.2 minutes per index task, faster than the 7.5-minute median. It is also verbose: 115M output tokens across the index against a 90M median. A low unit price still produces a larger bill when outputs run long, so measure output token volume in real use.

By reasoning setting, Ember scores 33.9 at about $0.116 and 35.8 at about $0.136. Claude Haiku 5.5 reaches 36.2 at about $0.411, but scores 34.3 at about $0.248 and 29.6 at about $0.159. GPT-6 Luna reaches 27.8 at about $0.152, its highest listed setting. The mapping of settings to named effort levels comes from the chart legend and was not confirmed.

The Muse family mixes hosted models and open-weights models, which are easy to confuse.

Ember is an API model, a separate line from the open-weights Glimmer 30B and Muse Spark Open. If you want to run something on your own GPU or Mac, Ember is not an option. Check the requirements in the two articles above instead.

Muse Ember is available in Muse Code, the Meta Model API, Google Cloud Model Garden and Oracle Cloud. On the Meta Model API you specify the model ID muse-ember-1.0 and call it through the Responses API. Meta's blog points to dev.meta.ai for trying it. The blog does not give endpoint URLs or SDK code, so this article does not either. Follow the official Meta Model API documentation at developer.meta.com/ai for the implementation. The general shape looks like this.

Meta suggests two patterns. One is delegating high-volume subtasks such as reading files, running tests and making targeted fixes. The other is pairing it with Muse Wonder, which plans the hardest problems while Ember handles the volume. With a 1M-token context you can load a whole codebase or document set in one call, but since prompts over 200K tokens triple the unit price, trimming to what is needed is usually wiser.

We could not confirm Claude Haiku 5.5 or GPT-6 Luna pricing from official pages, so those cells say to check the official source. The table covers what AA and Meta publish.

As a rule of thumb: for large volumes of agent calls centered on tool or computer use, Ember is a strong candidate. Where code mergeability matters most, Claude Haiku 5.5 is worth testing because it leads FrontierCode. For hard problems that need the highest ceiling, use a larger model such as Muse Spark 1.3 or Claude Fable 5.2 / Mythos 5.2 and route preprocessing and split-off work to Ember. Compare on success rate and total cost for your own tasks, not on unit price.

- Parameter count and open-weights status are not stated, so do not assume local or on-premises use
- Most benchmark figures come from Meta, and OSWorld 2.1, PerceptionBench, Agents' Last Exam and the Personal Superintelligence Index are Meta-internal
- Terminal-Bench 4.0 uses a different harness for each model, so conditions are not matched
- Prompts over 200K tokens triple the price of input, cached input and output
- AA rates the output as somewhat verbose, which can inflate real costs
- Competitor pricing was not verified here; check official pages before comparing
- Review the data-handling terms of your provider (Meta Model API, Google Cloud or Oracle Cloud) before sending business data