Playbooks · Coding
Muse Ember: Meta's cheap sub-agent model
Written November 10, 2026
Meta's smallest Muse model is built for the work an agent repeats all day: tool calls, coding subtasks, routing, and extraction. It is cheap by the token, until a request passes 200,000 input tokens and the whole request bills at three times the rate.
The problem
A builder routing sub-agent work to the cheapest model needs to know where the cheap price stops. Ember's headline rate looks like a tenth of a cent per thousand input tokens. That rate holds only while the request stays small. Past the threshold, the same request costs three times as much, and the threshold counts cached input and stored conversation history, not just fresh tokens.
The price, and the threshold
Meta's launch post puts Ember at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens, with a 1M-token context window and up to 512K output tokens, under the model ID muse-ember-1.0. Meta's pricing page adds the second tier:
| Request size | Input | Cached input | Output |
|---|---|---|---|
| Up to 200,000 input tokens | $0.10 | $0.004 | $0.50 |
| Over 200,000 input tokens | $0.30 | $0.012 | $1.50 |
All figures are US dollars per million tokens. The higher rates apply to the whole request, not just the part above the threshold. A request with 190,000 input tokens and 10,000 output tokens costs about $0.024; the same shape at 210,000 input tokens costs about $0.078, because the entire request moves to the higher tier.
What it is for
Ember is the smallest model Meta Superintelligence Labs sells, and the first model trained using Muse Wonder Full. Meta's own guidance pairs it with a larger model: Muse Wonder or Muse Spark plans, Ember handles the volume. The jobs Meta names are sub-agent tasks, tool calls, routing, extraction, and long document parsing. It is a reasoning model with no off switch: effort runs from minimal to max, and max is the default, so a call left at defaults thinks as hard as it can on every request. Reasoning tokens bill as output, so the effort setting is part of the price.
Ember is available on the Meta Model API in Standard mode only, and in Muse Code on pay-as-you-go and every subscription plan. It is also on Google Cloud and Oracle Cloud, and listed on OpenRouter, where an early build circulated as Crouton Alpha. There are no open weights.
How it scores
On Artificial Analysis' Intelligence Index, Ember scores 36 at maximum reasoning effort, level with Claude Haiku 5.5 at 36 and ahead of GPT-6 Luna at 28. Meta puts the cost at about $0.14 per index task, against $0.41 for Haiku, mostly because Ember writes fewer output tokens. On AutomationBench, Meta's table gives Ember 37.8%, against 13.9% for Haiku and 20.2% for Luna. On FrontierCode 2.0, the order flips: Haiku leads at 34.9%, Ember scores 27.2%.
Read Meta's table with care. Several rows are Meta's own evaluations, and the Artificial Analysis rows come from Artificial Analysis' November 9 run. Terminal-Bench ran each model in its own best harness, so it is not a like-for-like test. Nothing here is an independent run of ours.
When not to use it
Meta's docs send hard single-shot coding problems, knowledge-heavy questions, and visual coding to the bigger models. Ember is the model for bounded, repeatable steps with clear completion criteria, where a bigger model can check the result. A request that routinely carries a long conversation history will cross the 200,000-token line and lose the price advantage that justifies using Ember at all.
Limits
- Every benchmark figure on this page comes from Meta's launch post or from Artificial Analysis as reported there. No independent lab has reproduced these scores, and nothing here is a run of ours.
- The price tiers come from Meta's pricing page. The threshold counts cached input and stored conversation history toward the 200,000-token total.
- Ember is built for sub-agent work, not as a general-purpose model. Its scores on hard single-shot coding and knowledge questions trail the bigger models.
Related playbooks
- Run Muse Spark in another harness via the Model API
- DeepSeek's cheap coding model changed under the same name: retest Flash 1017 before you rely on it
- BotPsy 0.8B: an agent model that fits on-device
- Check whether your session is really on Muse Wonder or Wonder Full — and which route you are paying through
Sources
- Introducing Our Smallest, Most Efficient Model: Muse Ember — Meta
- Pricing and rate limits — Meta Model API
- Muse Ember — Meta
- Meta Muse Ember explained: specs, uses, and how to call it — eesel
- Meta Launches Muse Ember for High-Volume Agent Work — Zeniteq
Back to: Coding playbooks hub