Meta releases Muse Ember, a small model built for agent subtasks
Meta launched Muse Ember on 9 November 2026: a low-cost model with a 1M-token context, priced at $0.10 per million input tokens, for high-volume agent subtasks.
· 2 min read · By Efe Anidi
On this page +
Meta released Muse Ember on 9 November 2026, describing it as its smallest and most efficient model. Rather than competing on open-ended reasoning, Ember is built for the high-volume, repetitive work that AI agents run constantly: tool calls, coding subtasks, routing decisions, data extraction and parsing long documents. It ships under the model ID muse-ember-1.0 and is available immediately through Muse Code and the Meta Model API.
What Ember is for
Meta frames Ember as the execution half of a two-model pairing. For the hardest problems, Muse Wonder does the planning and Ember carries out the resulting steps at scale and at low cost. Meta also says Ember is the first model it has trained using Muse Wonder Full. It reads text, images, audio and video natively, holds a 1M-token context window, and can produce up to 512K tokens of output. It connects over the Responses API.
Pricing and availability
For prompts up to 200K tokens, Meta charges $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. Past 200K tokens, those rates rise to $0.30, $0.012 and $1.50 per million respectively. Ember is live now on Muse Code and the Meta Model API, and Meta says Muse Ember, Muse Spark and Muse Wonder are now also available on Google Cloud and Oracle Cloud. Meta names Box, Cline, Oracle, Cloudflare and You.com as early testers. Cloudflare’s chief technology officer, Dane Knecht, said Ember “ran up to 30% faster at the tail than the comparable model we tested.”
How it measures up, by Meta’s own numbers
Meta published benchmark comparisons against Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna, all run at each model’s maximum reasoning setting. On its figures, Ember scores higher than both rivals on AutomationBench-AA (37.8% versus 13.9% and 20.2%), Terminal-Bench 4.0 (40.9% versus 39.2% and 16.4%, each model run in its best published harness) and OSWorld 2.1 with partial credit (69.5% versus 68.9% and 50.5%). Haiku 5.5 came out ahead on FrontierCode 2.0-AA, scoring 34.9% against Ember’s 27.2%. These numbers come entirely from Meta’s own release and have not been checked by an outside benchmark group.
Why it matters for AI-native dev tooling
Agent systems built on a single large model tend to spend most of their tokens on routine steps: reading a tool’s output, extracting a field, deciding which function to call next. A cheaper, faster model built specifically for that layer, paired with a more capable planner, is a cost and latency lever that teams building agent pipelines and coding assistants can act on today, without redesigning the agent itself. Ember’s pricing and 1M-token context window make it plausible as the default “worker” model behind a larger planning model in that kind of setup.
What’s unconfirmed
Meta has not published Ember’s parameter count, model architecture or training data, and the announcement does not mention open weights. The safety section says Ember was evaluated under the Meta Superintelligence Scaling Framework, including chemical and biological, cybersecurity and loss-of-control risks, and stayed below the framework’s high-risk thresholds, but it gives no numeric results. Meta’s benchmark comparisons are self-reported and have not yet been reproduced independently.
Questions and answers
What is Muse Ember?
Muse Ember is a small AI model Meta released on 9 November 2026 as part of its Muse family. Meta built it for the high-volume, repeatable work that AI agents do, such as making tool calls, writing coding subtasks, routing requests, extracting data and parsing long documents, rather than for open-ended reasoning.
How much does Muse Ember cost and where can I use it?
For prompts up to 200K tokens, Meta prices it at $0.10 per million input tokens and $0.50 per million output tokens; longer prompts cost $0.30 and $1.50 per million. It's available now through Muse Code and the Meta Model API, and on Google Cloud and Oracle Cloud.
How does Muse Ember compare to other small models?
Meta's own benchmark comparisons, run at maximum reasoning effort, put Ember ahead of Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna on tests including AutomationBench-AA, Terminal-Bench 4.0 and OSWorld 2.1. Haiku 5.5 scored higher on one of the benchmarks Meta published, FrontierCode 2.0-AA. These are Meta's own published figures and have not been independently verified.
Can I download Muse Ember's weights?
Meta's announcement does not mention open weights. It offers Muse Ember as a hosted model through Muse Code, the Meta Model API, Google Cloud and Oracle Cloud.
Source:
