Meta* launches Muse Ember: an agent for code and documents at 14 cents per task

If you have deployed AI agents in real workflows, you know the main pain point: a capable model is expensive, while most of an agent's work is mundane. Read a document, run tests, change a couple of lines. Paying the same rate for this routine work as for complex reasoning seems odd, yet that is exactly what many people do. Meta* now has its own answer: the company has released Ember, the fourth model in the Muse family.

What we know about Ember

According to RB.RU, Ember is designed for typical agent tasks: analyzing documents, testing code, and making small code changes. The model is positioned as the lightest and cheapest in the Muse lineup. A single task costs 14 cents.

It is already available in two places: through the Meta* API (a programming interface that lets your system automatically send requests to the model and receive responses) and in Muse Code, the company's developer tool.

For now, that is almost everything that has been disclosed publicly. The source provides no details about the architecture, model size, context window, or benchmark results, so it is too early to make sweeping claims about quality.

What this looks like in practice

Imagine a typical day at a small team. In the morning, a batch of minor tickets lands in the repository: update a dependency, fix an error message, add a check for an empty value. Meanwhile, a lawyer or product manager needs to skim a dozen contracts or specifications and pull out the key terms.

For tasks like these, you do not need the most powerful model on the market. You need one that handles simple work reliably without breaking the bank when run a thousand times a day. The usual setup in these cases looks like this: a cheap model handles the stream of routine work, while an expensive one steps in only when a task goes beyond that scope — complex refactoring, a subtle bug, or an architectural decision. Judging by its description, Ember is aiming for precisely this workhorse role.

A price of 14 cents per task is useful because it is easy to build into your calculations. If an agent completes a hundred small tasks a day, that comes to about $14. Comparing that with the cost of human time takes only a moment.

What remains unclear

The main question is what exactly Meta* counts as a "task." One API call, one complete agent cycle involving several steps, or an average ticket? The answer determines whether the actual bill will be anywhere near the advertised 14 cents. For agents that loop, read files, and rerun tests, token usage can easily reach several times that of a single request.

The other unanswered questions are practical, too:

• How reliable is Ember at editing code? A cheap model that breaks tests costs more than an expensive one.
• How does it handle long documents and complex structures?
• What limits and terms apply to the API, and can the model be used in production without restrictions?
• How does Ember compare with similar lightweight models from other vendors on price and quality?

What developers and product managers should do

Do not switch everything to Ember at once. Pick one recurring workflow, such as running tests after small changes or extracting fields from documents, collect 50–100 real examples, and compare the results and total cost with what you use now. Measure the cost per successfully completed task, including retries and manual review, rather than the price per request.

The signal for founders is clear, too: major players are increasingly competing on the price of routine agent work as well as model capability. That is good news for anyone building products around agents: the cost of automation continues to fall.

But until independent tests are available, Ember remains a promise with an attractive price tag. Test it on your own data.

*Meta has been designated an extremist organization and is banned in Russia.