Meta's Mystery 'Crouton Alpha' Model Was Muse Ember All Along

Meta tested Muse Ember in public under the anonymous name Crouton Alpha, processing more than 5 trillion tokens before today's launch.

Jolly the cream-furred creature in a warm rustic bakery holding a small wooden bowl of golden croutons, oven glow behind, no text

Meta tested Muse Ember in public for a week before anyone knew what it was. The free, anonymous model listed as “Crouton Alpha” on OpenRouter since November 3 was an early version of Ember, Meta chief AI officer Alexandr Wang confirmed on Monday night, and the public test processed more than 5 trillion tokens before launch.

Wang said the team “fixed a bunch of the issues before launch” and thanked developers for the testing. OpenRouter’s listing for Crouton Alpha now states it “was developed and operated by Meta, revealed to be an early version of Muse Ember.” Wang’s post is here.

A Week of Guessing

Crouton Alpha appeared on OpenRouter on November 3 under the marketplace’s “stealth” label, which hides a model’s developer during a public preview. It was free to use, with a 1M-token context window and text, image, audio, video, and file input, and it quickly became one of the busiest models on the platform. Usage climbed from about 1.57 trillion tokens on November 5 to about 2.62 trillion the next day, according to OpenRouter’s public rankings as reported by Machine.net.

The clues pointed at Meta. Crouton Alpha shared the same six reasoning settings as Muse Wonder and Muse Spark 1.3, the same 1M-token context, and the same blocked-country list as Meta’s model service, and it failed on the same rare Japanese word that trips up Muse models. What no one could confirm was whether it was a Muse model at all, let alone which one.

The Test Behind the Launch

Ember launched earlier Monday as Meta’s smallest and cheapest model, built for the steps agents repeat all day: tool calls, coding subtasks, routing, and long document parsing. It costs $0.10 per million input tokens and $0.50 per million output tokens, and it matches Claude Haiku 5.5 on Artificial Analysis’s Intelligence Index at about a third of the cost per task.

Jolly the cream-furred creature cupping a small glowing ember in a cozy workshop at dusk, warm amber light, tools and shelves behind it, no text Earlier coverage Meta Releases Muse Ember, Its Smallest and Cheapest Model Yet

Ember is the first model trained using Muse Wonder Full, and the Crouton Alpha preview gave Meta a week of real developer traffic on an early build before the formal release. Ember is available now as muse-ember-1.0 through the Meta Model API and Muse Code, and on OpenRouter as Meta’s own listing.

Keep reading