Meta's Muse Ember ships at $0.10 per million input tokens

Meta Superintelligence Labs released Muse Ember on November 9, its smallest and most efficient model yet, at $0.10 per million input tokens.

The model id is muse-ember-1.0, and Meta built it for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing, according to its launch post. It is available now on the Meta Model API and inside Muse Code, Meta's coding agent, and accepts text, images, audio and video.

The cost per task carries the launch. Artificial Analysis puts Muse Ember at 36 on its Intelligence Index for $0.14 per task, the cheapest model to reach 36. Claude Haiku 5.5 (max) scores the same 36 for $0.41 per task, close to three times as much. Meta's own comparison puts sibling Muse Spark 1.3 one point higher, at 37, for $2.39 a task: a point for roughly 17 times the price.

Here's How Muse Ember's Pricing Compares

Prices are each maker's list rate; index scores are Artificial Analysis's at max reasoning. Meta's rate card adds a second tier: past 200,000 input tokens in one request, rates rise to $0.30 input, $0.012 cached input and $1.50 output per million, with cached tokens and carried-over conversation counted toward that line.

Muse Ember is the fourth model in Meta's Muse line. TechCrunch traces it through Muse Spark in April, open-weight Muse Glimmer in August and frontier Muse Wonder in October, with the consumer Muse Agent shipping in September. Meta says Ember is the first model it trained using Muse Wonder Full.

The price matters most for the loops a small video shop runs. Meta's rate card bills a call that moves a million input tokens and returns 200,000 at $0.20. The same call on Muse Spark 1.3, at $1.25 and $4.25 per million, costs $2.10. Run 5,000 of those a month and the bill is $1,000 against $10,500. A frozen brief cuts it further: Meta lists cached input at $0.004 per million, so re-sending a 50,000-token brief on 1,000 calls costs $0.20 instead of $5.00.

Video is a native input, the part a video shop can use. Meta's post says agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model. Meta's rate card lists one input line with no separate media meter, putting video on the same $0.10 as text.

One caveat sits under the headline figure. The 200,000-token line re-prices the whole request, not the tokens past it: VentureBeat works a request at 200,000 uncached input tokens with 10,000 billed output tokens out to 2.5 cents, then notes that one more input token takes the same request to about 7.5 cents. Meta also defaults Ember to its maximum reasoning effort, and reasoning tokens bill at the output rate. It lists no batch rate, where Anthropic halves Haiku 5.5 and OpenAI prices a batch tier for GPT-6 Luna.

Independent testing flags a capability weak spot. Artificial Analysis reports Ember's AA-Omniscience accuracy at 21.1%, the lowest of the similarly intelligent models it compared, because the model declines to answer 64% of the questions. That caution keeps its hallucination rate at 17.4%, about half of Haiku 5.5's.

Most of the benchmark rows in Meta's post, including Agents' Last Exam, the Personal Superintelligence Index, OSWorld 2.1 and PerceptionBench, are Meta's own internal evaluations. The index score and the cost-per-task figure come from Artificial Analysis. Meta did not disclose a parameter count.

For a creator paying per render, the test is narrower: whether $0.14 per task holds as an agent harness fires thousands of calls, or whether the 200,000-token line is where cheap stops being cheap.

Muse Ember is available now on the Meta Model API and Muse Code as muse-ember-1.0; it also runs on Oracle Cloud, and Google Cloud carries a pre-GA preview.