Meta launched Muse Ember on November 9. It's the smallest and cheapest model in the Muse family, built for the repetitive work AI agents grind through all day: tool calling, small coding tasks, routing requests, pulling information out of files, and reading long documents. Pricing is $0.10 for every million tokens you send in and $0.50 for every million it writes back.
What happened
Meta now has three Muse models, each with a clear job:
| Model | Role |
|---|---|
| Muse Wonder | The hardest problems. Works out the plan. |
| Muse Spark | The main model for coding and agent workflows |
| Muse Ember | High-volume small tasks, cheap enough to put in every product |
Meta's suggested setup is to let Wonder make the plan while Ember handles the bulk of the work. Meta says Ember is the first small model it trained using Muse Wonder Full.
Pricing and specs
| Item | Details |
|---|---|
| Model name | muse-ember-1.0 |
| Price (prompts up to 200K tokens) | Input $0.10, cache hit $0.004, output $0.50, all per million tokens |
| Price (over 200K tokens) | Input $0.30, cache hit $0.012, output $1.50 |
| Context window | 1M tokens, up to 512K output tokens |
| Input types | Text, images, audio, video |
| Where to use it | Meta Model API and Muse Code. Google Cloud and Oracle Cloud also offer it. |
| API | Supports the Responses API |
This release is API-only. There are no downloadable weights.
How it compares with similar models
Meta compared it with two other models that also focus on low cost, Claude Haiku 5.5 and GPT-6 Luna. All three were set to their highest reasoning level.
| Benchmark | Muse Ember | Haiku 5.5 (Anthropic) | Luna (OpenAI GPT-6) |
|---|---|---|---|
| AutomationBench (enterprise process automation) | 37.8 | 13.9 | 20.2 |
| Terminal-Bench 4.0 (command-line coding) | 40.9 | 39.2 | 16.4 |
| OSWorld 2.1 (operating a computer) | 69.5 | 68.9 | 50.5 |
| FrontierCode 2.0 (real engineering tasks) | 27.2 | 34.9 | 25.2 |
Meta says Ember matches Haiku's best scores at roughly a third of the cost. But on the harder real engineering tasks, Haiku still leads. Early feedback from partners centers on speed and on Ember working well as a subagent. Cline says it fits well as a subagent in workflows for writing code and analyzing docs, and Cloudflare found that on its slowest batch of requests it was about 30% faster than comparable models.
What this means if you build AI agents
If your AI agent runs hundreds or thousands of small steps a day, like batch-reading contracts and scoring them, sorting support tickets by rules, or running tests and fixing small bugs in a codebase, the model costs add up fast. Moving those steps to Ember and keeping your main model for the parts that need deep thinking could noticeably lower your bill.
The safest way to test it is to pick one repetitive task your agent does every day, run the same 20 samples through your current model and through Ember, and compare how many each got right and what each one cost.
Watch out
- The benchmark scores come from Meta. Some come from Artificial Analysis and some from Meta's internal tests. There's not much independent testing yet.
- It's not for hard puzzles. For complex architecture design or open-ended research, Meta itself recommends a bigger model.
- Long prompts cost three times as much. Once a prompt passes 200K tokens, the per-token price goes up across the board, so estimate costs before you process very long documents.
- No open weights. If you want to run it locally, it isn't an option.
Comments (0)
No comments yet. Start the conversation.