Meta’s Muse Ember shows up on OpenRouter at $0.10 per million input tokens

Meta has added a model called Muse Ember to OpenRouter, where Meta itself is the listed provider. The most concrete detail is the price: $0.10 per million input tokens and $0.50 per million output tokens.

That is nearly all the listing says. The page is an API guide rather than a model card, and it gives no description of what Muse Ember is for, no context length, no release date and no benchmarks. Meta has not published a separate announcement that we could find, so this article reports only what the listing shows.

What the listing shows and what it leaves out

According to the OpenRouter page, the model ID is meta/muse-ember and it is reached through the standard Chat Completions endpoint, POST https://openrouter.ai/api/v1/chat/completions. Only text is listed as a modality. The model accepts the usual request parameters, including tools, tool_choice, response_format, temperature, top_p, top_k, repetition_penalty, max_tokens and streaming. It also accepts reasoning and reasoning_effort, which suggests the model has some form of reasoning mode, though the page does not explain it.

Billing is pay-as-you-go with no OpenRouter markup, according to the page. The page also notes that EU in-region routing is not currently available, which matters for European companies with data-residency rules. Two further price entries are referenced on the listing, but their values were not visible in the text we could read, so there may be extra charges, for example for cached tokens, that are not covered here.

Anyone with an OpenRouter key can call the model now. The page gives the standard error codes, including 402 for insufficient credits and 429 for rate limiting.

Not established: Size, context and quality

The listing does not state parameter counts, architecture, context window or training data. It does not say whether Muse Ember is a small, fast sibling of Meta’s larger Muse models, a distilled version or something else, and it makes no performance claims. The name suggests a place in the Muse family, but the page itself does not say so. Any statement about how it compares with other models would be a guess, and there are no independent tests yet.

Because the price is the only hard number, it is the only basis for comparison. Output tokens cost five times as much as input tokens, a common ratio. For context, Tencent’s Hy4 appeared on OpenRouter this month at $0.417 per million input and $1.251 per million output, so Muse Ember’s listed rates are roughly a quarter and two fifths of those figures. Whether the models are comparable in ability is unknown.

A step down from Meta’s first paid API prices

Where this fits in Meta’s recent history is clearer than what the model is. In July, Meta launched a public preview of its Meta Model API with Muse Spark 1.1, priced at $1.25 per million input tokens and $4.25 per million output tokens, with a 1,048,576-token context window, according to Implicator. That report said the preview was limited to US developers, with a waitlist for new users, and that Meta was serving the model from its own platforms rather than through third-party marketplaces such as OpenRouter.

Measured against those figures, Muse Ember’s listed input price is one twelfth and its output price about one eighth of the Muse Spark 1.1 launch rates. That could point to a smaller or cheaper tier, but the listing does not say so, and the models may differ widely in capability. What the listing does show is that Meta is now serving a model through OpenRouter, a change from the July position described by Implicator. Meta’s open-weight releases and its API models have so far been separate lines; whether Muse Ember belongs to either is not stated.

For developers, the practical change is access. A Meta-served model is now reachable with one OpenRouter key and no separate Meta account or US-only waitlist, at least as the listing presents it. Teams that want to try it can test it cheaply, but should wait for Meta to publish specifications, a context length and benchmark results before moving any workload.