Meta released Muse Ember on November 9, which it describes as its smallest and most efficient model to date. It is built for the high-frequency grunt work that AI agents do: tool calling, coding subtasks, routing, information extraction and long-document parsing. The model natively handles text, images, audio and video, with a 1 million-token context window and up to 512,000 tokens of output per request.
Ember is available through the Meta Model API (model name muse-ember-1.0, with Responses API support) and Muse Code. It is also live on Google Cloud and Oracle Cloud, and developers can try it at dev.meta.ai. The announcement says nothing about open weights, so for now it appears to be an API-only model. Meta also noted that Ember is the first model trained with Muse Wonder Full.
Two-tier pricing
Ember is priced by prompt length, per million tokens:
| Item | Up to 200,000 tokens | Over 200,000 tokens |
|---|---|---|
| Input | $0.10 | $0.30 |
| Cached input | $0.004 | $0.012 |
| Output | $0.50 | $1.50 |
Its sibling Muse Spark 1.3 costs $1.25 for input and $4.25 for output. On the Artificial Analysis intelligence index the two score 36 and 37, a gap of one point, yet the cost of a single benchmark task is $0.14 versus $2.39, a difference of roughly 17x.
Benchmarks: some wins, some losses
Meta compared Ember at max reasoning effort against Claude Haiku 5.5 and GPT-6 Luna. The widest gap is on AutomationBench-AA, where Ember scores 37.8%, Haiku 5.5 scores 13.9% and GPT-6 Luna scores 20.2%. On the computer-use benchmark OSWorld 2.1, scored on a binary basis, Ember reaches 33.8% against 27.8% for Haiku 5.5. On Terminal-Bench 4.0, Ember scores 40.9%, just 1.7 percentage points above Haiku 5.5's 39.2%.
Coding is where it loses. On FrontierCode 2.0-AA, Ember scores 27.2% while Haiku 5.5 scores 34.9%. On GDP.pdf-AA it ties GPT-6 Luna at 22.8%. On the composite Artificial Analysis Index, Ember and Haiku 5.5 both score 36, and Luna scores 28.
The cost scatter plot gives finer numbers. Ember's two reasoning tiers cost $0.116 and $0.136 per task, with index scores of 33.9 and 35.8. Haiku 5.5 only reaches 36.2 at its $0.411 tier. At equal index scores, Ember's per-task spend is roughly one-third of Haiku 5.5's.
A rough daily bill for one subagent
Here is a back-of-the-envelope estimate using the announced prices. Assume a coding subagent reads 30,000 tokens per call and outputs 2,000 tokens, of which 20,000 input tokens are a cache-hit system prompt and code context:
- Uncached input of 10,000 tokens: about $0.001
- Cached input of 20,000 tokens: about $0.00008
- Output of 2,000 tokens: about $0.001
That comes to about $0.002 per call, or roughly $21 a day at 10,000 calls. This applies only to prices below the 200,000-token line; once long-document parsing crosses it, all three rates triple. The actual bill also depends on the thinking tokens generated at each reasoning tier, and the announcement does not say how those are billed separately.
Partner feedback centers on speed. You.com offered a concrete latency figure:
"with the lowest latency of any model we tested: a p95 of 2.2 seconds"
Cline said the model suits subagent work such as coding and document analysis. On safety, Meta says it evaluated the model under its Superintelligence Scaling Framework before release and found chemical and biological, cybersecurity and loss-of-control risks all below the high-risk threshold. All benchmark figures above were published by Meta, and independent retests have not yet appeared.
Sources: Meta official research blog, CocoLoop; the blog was used to verify the tiered pricing, benchmark scores, per-task costs and partner quotes, while the daily bill is an estimate based on the published unit prices.