Meta's Muse Ember undercuts the rest of its AI lineup at $0.10 per million tokens

Meta pushed Muse Ember live on the Meta Model API, billing it as the smallest reasoning model it has shipped. Input runs $0.10 per million tokens, output $0.50 per million. That makes it the bottom rung of the Muse lineup, a family of multimodal reasoning models from Meta Superintelligence Labs built for tool use and mixed inputs like text, images and audio.

The price gap to its own siblings is steep. Muse Spark 1.3 sits at roughly $1.25 per million input tokens and $4.25 per million output tokens, meaning it costs about 12.5 times Ember on input and 8.5 times on output. Muse Spark and Muse Wonder, the bigger models above Ember, are aimed at heavier workloads where capability beats cost per query.

Ember's rates land almost exactly on the Meta Model API's discounted Contributor tier. Developers on that tier save real money, but they sign off on Meta using their interaction data to train future models. Privacy-sensitive teams and companies in regulated industries may find that trade hard to swallow, and it hands Meta a steady stream of real-world data if enough Ember users opt in.

What is missing is a scorecard. No independent benchmarks or technical specifications for Ember have surfaced publicly, leaving it unclear how it stacks up against Spark, Wonder or rival small models. The Muse family only started in April 2026 with Muse Spark, so Ember gives the lineup a cut-price entry point about six months in.

Ember lands as Meta's smallest reasoning model as of October 2026, arriving after a restructuring of its AI efforts that began around 2025 and included the creation of Superintelligence Labs. That push has tilted Meta toward proprietary models and tiered access — and at a dime per million input tokens, reasoning capability opens up to teams that higher-end pricing had shut out.