Meta* Releases Muse Ember: 1M Tokens of Context and $0.10 per Million Input Tokens

If you have built agents on large models, you know what it looks like on the bill. An agent doesn't make a single request. It goes in circles: it calls a tool, reads the response, thinks, then calls the next one. At every step the entire accumulated context is sent to the model, and the cost grows faster than it seems at the start. This is exactly the pain Meta* is trying to address with Muse Ember.

*Meta is an organization recognized as extremist, and its activities are banned in the Russian Federation.

What the model is and what it was built for

Muse Ember is a compact model designed for cyclical agent tasks. Meta* lists four use cases: tool calling, parsing bulky documents, routing, and writing code snippets. These are the workhorses of agent systems. What matters here is not genius on olympiad-level problems, but predictability, speed and a low price per step.

The specs are as follows:

• context window: 1M tokens;
• response length: up to 512K tokens;
• multimodality: audio, video and images as input;
• API identifier: muse-ember-1.0;
• availability: the Meta* API, Google Cloud and Oracle Cloud.

Price: 10 cents per million input tokens and 50 cents per million output tokens.

What this means in practice

Let's run the numbers on a hypothetical example. Suppose your agent sends 50K tokens of context to the model at each step and receives about 1,000 tokens in response. At Ember's rates, that is roughly half a cent for input and one-twentieth of a cent for output, or about $0.0055 per step. A hundred steps in a complex task would cost around 55 cents. The calculation is ours and purely illustrative, but the order of magnitude is clear: a cheap model lets you avoid skimping on the number of iterations and re-checks.

The second practical advantage is multimodality without workarounds. The agent can read an interface screenshot or process a meeting recording on its own, and you don't need to plug in a separate speech recognition or vision model. Fewer components in the pipeline means fewer places where something can break and less latency between services.

The third use case is routing. A compact model is convenient to put at the front of a system: it classifies the request, decides whether a heavy tool or an expensive model is needed, and only then passes the task along. That way, expensive calls are reserved for genuinely hard cases.

Where to be more careful

The main caveat: the claim that Muse Ember beats Claude Haiku 5.5 and GPT-6 Luna on price-to-intelligence is based on Meta*'s own benchmarks. This is not an independent comparison. The "price to intelligence" metric is easy to tune to a favorable set of tests, and real-world agent tasks depend heavily on your toolset, prompt format and the quality of your function-calling schemas.

Second, long context. A million-token window doesn't guarantee that the model will find a needed detail equally well at the beginning and at the end of a document. And 512K tokens of output is an upper limit of capability, not a promise of quality across the full length. You need your own tests on your own data.

Third, availability in two clouds, Google Cloud and Oracle Cloud, is convenient for companies that already have contracts and billing there. But region and quota restrictions need to be checked separately, and Russian teams also face legal nuances in working with products of a company whose activities are banned in Russia.

What to do

1. Take 50–100 real traces from your agent that include tool calls and long documents.
2. Run them through muse-ember-1.0 and through the model you currently use.
3. Compare not only accuracy, but also the cost of a full run, the number of function-call errors and response time.
4. If the results are close, try putting Ember in the role of router or executor of simple steps, leaving the expensive model for planning.

The model is interesting first and foremost as one more source of price pressure in the segment of cheap agent models. The numbers in the report are worth believing only until you have checked them on your own tasks.