Muse Ember is really cheap. Meta has launched it at $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004. It’s available through Muse Code and the Meta Model API under muse-ember-1.0. 1
The price caught my attention straight away. In our post, I called it about three times cheaper than Haiku 5.5 per task for the same Artificial Analysis score. The detailed numbers are close rather than identical: 35.8 for Ember Max against 36.2 for Haiku Max, which both round to 36 in Meta’s summary. 1 2
| Model | AA Intelligence Index | USD per AA task |
|---|---|---|
| Muse Ember | 35.8 | $0.136 |
| Claude Haiku 5.5 | 36.2 | $0.411 |
| GPT-6 Luna | 27.8 | $0.152 |
That puts Ember at roughly a third of Haiku’s task cost in this comparison, while scoring above Luna Max for slightly less money. Meta estimates Ember’s task cost using its announced pricing; these aren’t Lumina test results. Task cost also depends on how many tokens a model uses, so the per-million-token price is only part of the bill.
This is where a small model can be useful. An agent can spend a lot of a session reading files, extracting information, calling tools and making small fixes. If a cheaper model handles those jobs well, the savings add up without asking it to solve the hardest part of every task.
Meta gives Ember a 1M-token context window and up to 512K output tokens, with native text, image, audio and video understanding. It supports the Responses API, and Meta says it is the first model trained using Muse Wonder Full. 1
The results aren’t a clean sweep. Meta’s table puts Ember ahead on Terminal-Bench 4.0 at 40.9%, against 39.2% for Haiku and 16.4% for Luna, but each uses a different best published harness. On FrontierCode 2.0-AA, Haiku leads at 34.9% against Ember’s 27.2%. I’d keep both of those details in mind before treating it as a replacement for every coding job. 1
For me, the appeal is getting useful work done at a price that makes it easy to keep going. We’ve already covered Muse Wonder at the larger end of Meta’s lineup. Ember gives us something very different to try, and the published price-performance comparison makes it a welcome addition.
Sources & references
- Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember ↗
9 November 2026. Official announcement read in the browser. Pricing, availability, capabilities and publisher-presented benchmark comparisons. Artificial Analysis v5 results are reproduced by Meta; internal evaluations and harness differences are identified in its methodology footnote.
- Lumina: Muse Ember’s price and Haiku comparison ↗
9 November 2026, 17:59 UTC. Our reaction to the published cost-per-task comparison, rather than a claim of firsthand Ember testing.
- Lumina: Muse Ember release finding ↗
9 November 2026. Our original release post and supplied findings card. The post is also quoted in the later price comparison.
