Meta releases Muse Ember: input tokens start at $0.10 per million

When an agent reads several files, selects a tool, and runs a test to make a small change, every intermediate step costs money. With thousands of such steps, model selection becomes a question of product economics. Meta released Muse Ember specifically for this work: tool calls, data extraction, document parsing, and small coding tasks.

A practical scenario is straightforward: one model draws up a plan for a fix, while Ember gets a discrete assignment—to find the right function or prepare a small patch. For developers, this is a reason to test whether they can cut spending on repetitive operations without increasing errors.

What lies behind cheap tokens

In the supplied analysis by Simon Willison, Ember, GPT-6 Luna, and Claude Haiku 5.5 have identical base prices: $0.10 per million input tokens and $0.50 per million output tokens. For a short request, the price list alone does not help identify a winner.

The difference emerges as context grows. Ember's threshold is 200,000 tokens: above that, input costs $0.30 per million and output costs $1.50. Luna's increase starts after 272,000 tokens, while Haiku's starts after 100,000. All three models support context windows of up to one million tokens, but charge for them under different rules.

Take a hypothetical request with 250,000 input tokens and 10,000 output tokens, without caching. According to the table provided, Ember would cost $0.09, Luna $0.03, and Haiku $0.15. This is a calculation for a specified volume, not a forecast of actual spending: the models may use different numbers of tokens for the same assignment.

Loading an entire repository into a single request is therefore technically possible, but its economic benefits need a separate calculation. Sometimes it is cheaper to select the relevant files first and send only those to the model.

Why reasoning settings change the comparison

Ember has six reasoning levels, from minimal to max. A higher setting may consume more tokens, so identical rates no longer mean an identical cost per result. In a product, the entire process matters: the first response, validation, a retry, and potentially handing a difficult case to another model.

In the charts in Willison's article, Ember ranks among the models offering the most favorable balance of performance and cost per task. But that comparison uses a particular set of tests. Applying it to your own service without testing would be premature.

There is also an important detail in the original source: Meta labels Ember's cost in the comparison chart as an estimate based on proposed pricing. In addition, on FrontierCode 2.0-AA, Ember scores 27.2%, compared with Haiku's 34.9% at its maximum setting. Even within the results presented, the leader depends on the task. These caveats appear in Meta's announcement.

Willison himself tested the reasoning levels with a familiar assignment: drawing a pelican riding a bicycle in SVG. The high setting did not return a complete SVG, and increasing the level did not consistently improve the image. It is a standalone experiment, but a useful reminder: the maximum setting does not guarantee the best response to every assignment.

What to check before integration

Start with one recurring stage in your agent's workflow. For example, extracting fields from documents or selecting a tool based on a user's request. Collect examples with verifiable correct results, including long inputs and ambiguous cases.

Run several reasoning levels and calculate the cost of an accepted result. Record latency, formatting errors, repeat calls, and the share of tasks that had to be passed to a more expensive model alongside it. For code, add test pass rates and the amount of manual correction required.

For a product manager, this test answers a specific question: how much it costs to handle a request or prepare a change at the required quality. For a founder, it provides a basis for recalculating a feature's unit cost. Ember deserves a place in that comparison; the results from your own workflow should determine whether to switch.