Meta's Muse Ember Goes After the Cheap-Model Market

Meta released Muse Ember on Monday, the newest and smallest member of its Muse family, and the pitch is almost entirely about price. Meta Superintelligence Labs calls it the lab's lightest and most cost-efficient model to date, priced at ten cents per million input tokens. As TechCrunch reports, that undercuts Anthropic's Haiku and OpenAI's Luna, and sits in the same bracket as open-source options like DeepSeek V4.1 Flash and GLM 5.3 Flash.

Ember is not trying to be a frontier brain. Meta describes it as built for "complex reasoning execution," which in practice means the repetitive machinery of agent work: tool calls, small coding subtasks, routing, data extraction and chewing through long documents. The intended setup pairs it with the larger Muse Wonder, released last month. Wonder lays out the plan and the rules, and Ember runs them at scale. Meta's own phrasing captures the division of labour neatly: "Muse Wonder plans, Muse Ember handles the volume."

A multimodal workhorse

Despite the budget framing, Ember is not text-only. Meta trained it on text, images and audio, so it can read screenshots, charts and scanned documents, and process voice memos, meeting recordings and video without handing off to a separate model. For a customer-support operation fielding tens of thousands of cases, that breadth matters more than a few points on a reasoning leaderboard.

On the numbers, Meta leans on Artificial Analysis's Intelligence Index to argue Ember is both cheaper and smarter than Haiku and Luna, and early coverage from aggregators reports it beating Claude Haiku 5.5 on an automation benchmark at roughly a third of the cost. Those are Meta's framings on Meta's chosen chart, so independent testing will tell the real story.

Where it sits in the Muse line

Ember is the fourth model Meta has shipped under the Muse banner this year, after the mid-weight Spark in April, the open-weight Glimmer in August and the frontier-class Wonder in October. Meta has not said whether Ember's weights are open, which leaves a gap given Glimmer was explicitly the family's open release. The broader pattern is clear enough: the frontier labs are no longer competing only on raw intelligence but on who can run reliable agent work for the least money. Meta's answer is to split the job in two and make the busy half cheap. It is the same logic now reshaping the model market from both ends at once.