Meta's smallest model is the one we can't run

Meta released Muse Ember today, and calls it its smallest and most efficient model ^[1]. It's built for the work agents repeat all day, the sub-agent tasks, tool calls, routing and the same job run over thousands of files or photos, while a bigger model does the planning ^[2]. We've listed it, and I think it's good for the industry. It's also the smallest model Meta makes, and the only way to use it is to rent it.

Is it available?

It's available as a service. It's on Meta Model API as muse-ember-1.0 and in Muse Code, on pay-as-you-go and every plan, at $0.10 per million input tokens and $0.50 per million output, with a 1M-token context, and text, images, video, audio and PDF going in ^{[2] [3]}. Meta even tells you how to put it in Claude Code, as the model behind the haiku alias and your sub-agents ^[4].

There's nothing to download, though. Meta's models page lists its hosted models and its open-weight ones, and Muse Ember is a hosted one only, with no weights and no parameter count ^[5].

Why it's good

Most of what an agent does isn't hard. It's the fiftieth file to classify, the hundredth tool call in a loop, the routing decision before the real work starts, and paying a frontier price for that is an easy way to burn money on agents. Meta's own worked example sorts 10,000 photos with Muse Wonder writing the pipeline and Muse Ember running it, for $1.33 in all, against about $20 for the same requests at Muse Wonder's prices ^[6]. That split, a big model planning and a small one doing the volume, is where agents are going, and every lab making its small model better and cheaper makes it cheaper for everyone.

A third of the cost of what?

The launch thread says Muse Ember gets more intelligence per dollar than Haiku, Luna and all other flash models, at a third of the cost or less. Meta's announcement makes a narrower claim than that. It compares it with Claude Haiku 5.5 and GPT-6 Luna, and says it matches Haiku's best score at about a third of the cost ^[1]. I couldn't find the "all other flash models" part anywhere on Meta's pages.

And it's a third of the cost per task, not per token. Muse Ember, Claude Haiku 5.5 and GPT-6 Luna all list $0.10 per million input tokens and $0.50 per million output ^{[3] [7] [8]}. The gap is in how many tokens each one spends to answer, measured by Meta on the Artificial Analysis tasks, where it puts Muse Ember at about $0.136 a task and Haiku at about $0.411 for scores of 35.8 and 36.2 ^[1].

That's a real advantage if it holds up, since fewer tokens is also less waiting. But it's Meta's measurement, several of the other rows in its table are Meta's internal benchmarks, and Haiku still wins on FrontierCode ^[1]. Two things on the pricing page change the sum too. Muse Ember defaults to the most reasoning it can do, so a routing call left on the default thinks as hard as it can ^[9], and a request over 200K input tokens is billed at three times the rate, for the whole request ^[3].

The one we can't have

Meta does publish open weights. Muse Spark Open is an open-weight version of Muse Spark, with 816B total and 42B active parameters ^[5], and Muse Glimmer is a model of about 30B parameters under Apache 2.0 that runs on your own hardware (listed here).

But an 816B model needs a datacenter, and Muse Glimmer is distilled from Muse Spark, not the model Meta built for this work ^[5]. Meta calls Muse Ember its smallest model and doesn't say how big it is, so I can't tell you it would fit on your desk. A small model doing the same task ten thousand times is what I'd want on my own machine though, next to my own files, with no meter running, and it's the first thing I'd put in our calculator. For now the only way to use it is to rent it, on an API with no status page with readable history and no deprecation policy that we could find (Meta Model API).

I'd like to be wrong about this, and for the weights to follow the way they did for Muse Spark. Nothing on Meta's pages says they will.

References

Read on 9 November 2026.

1. [1]Meta AI Research, Introducing our smallest, most efficient model, Muse Ember, 9 November 2026. Source for Meta calling it its smallest and most efficient model, the comparison with Claude Haiku 5.5 and GPT-6 Luna, matching Haiku's best score at about a third of the cost, the cost per task and scores ($0.136 and 35.8 against $0.411 and 36.2), Haiku leading on FrontierCode 2.0, and the internal benchmarks in the table. research.meta.ai
2. [2]Meta, Muse Ember, Meta Model API documentation. Source for what it's built for, the model ID, inputs, the 1,048,576-token context, and availability on Model API and in Muse Code on pay-as-you-go and every plan. dev.meta.ai
3. [3]Meta, Pricing and rate limits, Meta Model API documentation. Source for $0.10 input and $0.50 output per million tokens up to 200K input tokens, and the long-context rates of $0.30 and $1.50 for the whole request above that. dev.meta.ai
4. [4]Meta, Choose a model, Meta Model API documentation. Source for mapping Claude Code's haiku alias and sub-agent model to muse-ember-1.0. dev.meta.ai
5. [5]Meta, Models, Meta Model API documentation. Source for Muse Ember being one of the hosted models, Muse Spark Open being the open-weight release of Muse Spark with 816B parameters and 42B active, and Muse Glimmer being distilled from Muse Spark. dev.meta.ai
6. [6]Meta, the Muse Ember launch post, Meta for Developers blog, 9 November 2026. Source for the 10,000-photo example costing $1.33, with about $20 for the same requests at Muse Wonder prices. dev.meta.ai
7. [7]Anchor Terminal, the Anthropic API listing. Source for Claude Haiku 5.5's list price of $0.10 input and $0.50 output per million tokens, up to 100,000 prompt tokens. anchorterminal.com
8. [8]Anchor Terminal, the OpenAI API listing. Source for GPT-6 Luna's list price of $0.10 input and $0.50 output per million tokens. anchorterminal.com
9. [9]Meta, Reasoning, Meta Model API documentation. Source for Muse Ember's reasoning effort defaulting to max, and max using more reasoning tokens at the same per-token price. dev.meta.ai