Meta’s launch artwork for Muse Ember. Image: Meta
Meta has released Muse Ember, the smallest and cheapest model in its Muse family, built for the repetitive work AI agents do all day: calling tools, running coding subtasks, routing requests and pulling data out of long documents. Meta says it matches the best score of Anthropic’s Claude Haiku 5.5 on Artificial Analysis’s independent benchmark index at about a third of the cost per task.
It went live on November 9 on the Meta Model API and in Muse Code, Meta’s coding agent, under the name muse-ember-1.0, and is also listed on OpenRouter.
Same price per token as Haiku, fewer tokens per job
Muse Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. That is exactly what Anthropic charges for Claude Haiku 5.5 on prompts up to 100,000 tokens, and what OpenAI charges for GPT-6 Luna, going by OpenRouter’s listings. So Meta’s savings claim rests on how few tokens Ember uses to finish a job, not on a lower rate.
Meta’s figures come from version 5 of the Artificial Analysis Intelligence Index, published the same day. At its highest reasoning setting Ember scores 35.8 at about $0.14 a task. Haiku 5.5 reaches 36.2 at its maximum setting, but at $0.41 a task. GPT-6 Luna tops out at 27.8 for about $0.15.
Ember can take in a million tokens at once and write up to 512,000 in reply. It reads text, images, audio and video, so an agent can work from screenshots or meeting recordings without a second model.
Where it wins, and where it doesn’t
With all three models at maximum reasoning effort, Ember leads on most of the tests Meta picked:
- Business workflows: 37.8% on AutomationBench-AA, against 20.2% for Luna and 13.9% for Haiku.
- Terminal coding: 40.9% on Terminal-Bench 4.0, just ahead of Haiku’s 39.2%, with Luna on 16.4%.
- Computer use: 69.5% on OSWorld 2.1 with partial credit, against 68.9% for Haiku and 50.5% for Luna.
- Coding agents: 46 on Artificial Analysis’s Coding Agent Index, against 37 for Haiku and 31 for Luna.
It loses clearly on FrontierCode 2.0-AA, which judges real engineering work on whether it could be merged: Haiku scores 34.9 to Ember’s 27.2. On GDP.pdf-AA, professional work over long documents, Ember ties Luna at 22.8. Some of the strongest numbers, including OSWorld, Agents’ Last Exam and a “Personal Superintelligence Index” where Ember leads 46.8 to 37.0, are Meta’s own internal tests and haven’t been checked by anyone else.
Where it fits in the Muse line-up
When Meta launched Muse Wonder in October, it said it would use its biggest models to build smaller ones for everyone. Ember is the first: Meta says it was trained using Muse Wonder Full. The company now pitches three tiers: Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Ember for the same kind of work repeated at scale.
Against Muse Spark 1.3, the latest version of the model behind Meta AI, Ember gives up a single point on the Intelligence Index (36 to 37) but costs $0.14 a task instead of $2.39. Spark lists at $1.25 per million input tokens and $4.25 per million output tokens.
Unlike Muse Glimmer, Meta’s small Apache-licensed model, or Muse Spark Open, Ember can’t be downloaded: it runs only through Meta’s API and services that resell it. Meta says it stayed below its safety framework’s high-risk thresholds for chemical, biological, cyber and loss-of-control risks.
‘Purpose-built for subagent work’
Meta lined up five early testers: Box, Cline, Oracle, Cloudflare and You.com. Cline’s founder sees Ember as the cheap helper a bigger coding agent hands jobs to:
Muse Ember is purpose-built for, and excels at, subagent work like coding and document analysis.
Saoud Rizwan, founder, Cline
Cloudflare’s technology chief pointed to speed:
What stood out in our testing was speed. Muse Ember ran up to 30% faster at the tail than the comparable model we tested.
Dane Knecht, CTO, Cloudflare
You.com says Ember finished every coding task it set with a 95th-percentile response time of 2.2 seconds, which it puts at about four to 13 times faster than leading open-weight models. Those are partner claims, not independent tests.
Meta doesn’t list countries. As with Muse Wonder, availability follows the Meta Model API and Muse Code, which don’t yet cover every market; OpenRouter is another route in.
Why it matters
Agents now make hundreds of small model calls for every task a person sets them, so the cheap model doing the legwork matters as much as the flagship. If Artificial Analysis’s numbers hold, Meta has a model that does Haiku-level work on fewer tokens, putting pressure on Anthropic and OpenAI where they compete hardest on price.
Sources: Meta: Introducing our smallest, most efficient model, Muse Ember (November 9, 2026); Meta: Muse Ember evaluation report; AI at Meta on X; OpenRouter: Muse Ember.
Latest Meta news
