Meta's Muse Ember Is Its Smallest Model Yet — Haiku-Level Scores at a Third of the Cost
Photo: Hippopx (public domain). Illustrative image.
What happened. Meta released Muse Ember on Monday 9 November 2026, the smallest and cheapest model in its Muse family. Announced by Mark Zuckerberg on X as delivering “higher intelligence per dollar than others in its class”, Ember is built for the work AI agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing. It is live now on the Meta Model API and in Muse Code, Meta’s coding agent, as muse-ember-1.0, and is also listed on OpenRouter. Meta calls it the first model trained using Muse Wonder Full, its largest model, and pitches the family as a team: “Muse Wonder plans, Muse Ember handles the volume.”
The price is the point. Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. It takes text, images, audio and video as input, with a 1-million-token context window — enough to hold a full codebase or document set in one call — and can return up to 512,000 tokens in a single response. It supports the Responses API. Meta does not disclose a parameter count, and the weights are not released: this is an API-only model, not an open download.
The comparison Meta wants you to see. On Artificial Analysis’s Intelligence Index, Meta says Ember scores 36 — the same as Claude Haiku 5.5 at maximum reasoning — at $0.14 per task against Haiku’s $0.41, about a third of the cost. It also says Ember beats OpenAI’s GPT-6 Luna at every reasoning setting, and that no model on the full Artificial Analysis leaderboard scores higher for less per task. On AutomationBench, which tests business workflows, Ember scored 37.8%, nearly three times Haiku 5.5’s 13.9%. Against its own bigger sibling, Muse Spark 1.3, Ember scores 36 to Spark’s 37 on the same index — but at $0.14 per task against $2.39. All of these figures are Meta’s or Artificial Analysis’s, from launch-day evaluations; none has been independently replicated, and the cost-per-task comparisons use Meta’s list pricing.
The catch. Small models have weak spots, and Ember’s is knowledge: Artificial Analysis says it declines 64% of trivia-style questions. That is the trade-off in the design — Ember is built to execute repetitive tasks cheaply, not to answer questions from memory. Meta’s own framing concedes the point: Ember is the volume worker under Wonder, the planner, and Spark, the coder. Used as a subagent handed well-specified jobs, the knowledge gap matters less; used as a general assistant, it would show.
Who is using it already. Meta named early-testing accounts from Cloudflare and You.com, and early partners including Box, Cline and Oracle. Ember, Muse Spark and Muse Wonder are now available on Google Cloud and Oracle Cloud as well as Meta’s own API. Meta says Ember was evaluated under its Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks, though the evaluation report’s details were not the focus of the launch announcement.
Why it matters. The frontier-model race of the past two years was about the biggest, smartest model. Ember is the opposite bet: that the money in AI agents is in the millions of small, repetitive calls underneath the flagship — and that the winner will be whoever does those calls cheapest without breaking them. Anthropic cut Fable’s price 20% on the same day, and the two launches together say the same thing: the price war has moved from the flagship to the whole stack. For developers building agents, the question is no longer which model is smartest, but which combination of planner and worker does the job for the least money — and Ember is Meta’s answer for the worker half of that pair.
Sources
- Meta AI Research — “Introducing Our Smallest, Most Efficient Model: Muse Ember” (9 November 2026): https://research.meta.ai/blog/muse-ember
- Unite.AI — “Meta Releases Muse Ember for Lower-Cost Agent and Coding Work” (9 November 2026): https://www.unite.ai/meta-releases-muse-ember-for-lower-cost-agent-and-coding-work/
- Digital Trends — “Meta launches Muse Ember, a cheaper AI model built for repetitive tasks” (10 November 2026): https://www.digitaltrends.com/computing/meta-launches-muse-ember-a-cheaper-ai-model-built-for-repetitive-tasks/
- CellCog — “Muse Ember: Meta’s Smallest Model, Price and Benchmarks” (9 November 2026): https://cellcog.ai/blog/muse-ember/
- FreshSynapse — “Anthropic Launches Claude Fable 5.2 and Mythos 5.2 at Lower Prices — Same Model, Two Safety Levels”: https://freshsynapse.com/articles/anthropic-claude-fable-mythos-5-2-2026-11-10/
