The short answer
● All three cost $0.10 — Muse Ember, Claude Haiku 5.5 and gpt-6-luna are identical on input price.
● Haiku has a cliff — Past 100,000 tokens it jumps to $0.50 input and $2.50 output.
● Ember takes audio — Text, images and audio. The other two do not match that at this price.
Meta priced Muse Ember at $0.10 per million input tokens on Monday. That is exactly what Anthropic charges for Claude Haiku 5.5 and exactly what OpenAI charges for gpt-6-luna. When three labs land on the same number, price stops being a reason to choose and everything else starts mattering more.
|
Muse Ember |
Claude Haiku 5.5 |
GPT-6 Luna |
| Input / M tokens |
$0.10 |
$0.10 (under 100K) |
$0.10 |
| Output / M tokens |
Not published |
$0.50 (under 100K) |
$0.50 |
| Long-prompt pricing |
Not published |
$0.50 / $2.50 over 100K |
Flat |
| Modalities |
Text, images, audio |
Text, images |
Text, images |
| Batch discount |
Not published |
$0.05 / $0.25 |
Available |
| Released |
9 November 2026 |
October 2026 |
2026 |
| Built to pair with |
Muse Wonder |
Sonnet 5.5 / Opus 5.5 |
GPT-6 family |
What actually separates them
Prompt length. Claude Haiku 5.5 is $0.10 only while the whole prompt stays under 100,000 tokens, and Anthropic counts cache reads and cache writes toward that total. Cross it and input becomes $0.50 and output $2.50 — a fivefold jump on the same request. If you are feeding long documents, Haiku is not a $0.10 model.
Audio. Ember handles text, images and audio. At this price point that is unusual, and it is the clearest single reason to pick it over the other two.
Where your users already are. gpt-6-luna sits inside the OpenAI stack that a lot of teams already bill through. Adding a Meta endpoint to save nothing is work for no gain.
Which one to use
- High-volume short requests → any of them. Genuinely. Pick the one whose SDK you already have.
- Long documents or large caches → not Haiku 5.5, unless you have modelled the 100K cliff.
- Anything involving audio → Muse Ember.
- Already on OpenAI billing → gpt-6-luna. The switching cost outweighs a tie.
- You need a planner as well as an executor → Ember with Muse Wonder, or Haiku with Sonnet 5.5.
The gap in this comparison
Meta has not published Ember’s output token price, its batch rate or its context limit. Input price alone does not decide cost — output is typically five times input on this tier, and it is the number that moves a bill. Until Meta publishes it, treat Ember’s total cost as unknown rather than matched.