Meta's cheapest AI model yet costs 10 cents per million input tokens. On the benchmark that matters for AI agents, it ties Claude Haiku 5.5 — and Meta says each task costs about a third as much. Here is the comparison.
What Meta launched
On 9 November 2026, Meta introduced Muse Ember, the smallest and cheapest model in its Muse family, available now as muse-ember-1.0 on the Meta Model API and inside Muse Code.
Meta built Ember for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. It handles text, images, audio, video and documents, with a 1,048,576-token context window and up to 512,000 output tokens. Its weights are private, so you cannot run it on your own hardware.
The price, in plain numbers
The headline price is $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004. But those rates apply only up to 200,000 input tokens; above that, both triple, to $0.30 and $1.50.
| Muse Ember | Claude Haiku 5.5 | GPT-6 Luna | |
|---|---|---|---|
| Input price | $0.10 per million | — | — |
| Output price | $0.50 per million | — | — |
| Intelligence Index score | 36 | 36.2 | 28 |
| Cost per task | About $0.14 | About $0.41 | About $0.15 |
Index scores and per-task costs come from Artificial Analysis, using figures Meta published on launch day. Per-task costs are estimates at proposed pricing, not measured bills.
The benchmark scores
Artificial Analysis ran all three through its Intelligence Index, version 5, on launch day. At maximum reasoning effort, Ember scored 36, effectively tied with Haiku 5.5 at 36.2 and well ahead of GPT-6 Luna at 28.
On cost, Ember averaged about $0.14 per task, against $0.41 for Haiku and $0.15 for Luna. Meta claims Ember delivers more intelligence per dollar than Haiku 5.5 or GPT-6 Luna at any reasoning setting, and matches Haiku's best score at about one-third of the cost. That is Meta's claim, not an independently verified fact.
What each model is for
All three sit in the same lightweight class: small models for high-volume work inside agents, not flagship models for the hardest single questions.
Ember is the volume play. If your agent makes thousands of small calls a day, the per-task cost decides your bill. At about $0.14 per task, it is the cheapest of the three, and close enough to Haiku that the saving is hard to ignore.
Haiku 5.5 is the capability match. It scored marginally higher, and it lives inside Anthropic's ecosystem: the Claude Platform, Claude Code, and the AWS, Google Cloud and Azure marketplaces. If your stack is built on Claude, switching providers has its own cost.
GPT-6 Luna is the OpenAI-stack option. It scored lowest on the index, but its per-task cost is nearly the same as Ember's. For teams already on OpenAI, Luna may be the simpler choice.
The catches
First, the long-context pricing. If your agent regularly reads documents over 200,000 tokens, the headline rate does not apply, and the tripled rate changes the maths. Check your actual request sizes first.
Second, these are launch-day numbers: per-task figures are estimates at proposed pricing, and the scores come from a single evaluation run. Real workloads will differ.
Which one should you pick?
If cost per task is your main constraint, Ember is the strongest choice on the numbers published so far. If you are deep in Anthropic's ecosystem and want the highest small-model score, Haiku 5.5 is the safer pick. If your team runs on OpenAI, Luna keeps everything in one place at a similar per-task cost.
Key takeaways
- Meta launched Muse Ember on 9 November 2026 at $0.10/$0.50 per million input/output tokens.
- On the Artificial Analysis Intelligence Index, Ember scored 36, tied with Haiku 5.5 at 36.2 and ahead of GPT-6 Luna at 28.
- Per-task cost was about $0.14 for Ember, $0.41 for Haiku and $0.15 for Luna.
- The headline price applies only up to 200,000 input tokens; above that, rates triple.
- Ember has a 1-million-token context window and up to 512,000 output tokens.
- Ember's weights are private, so it cannot be self-hosted.
Conclusion
Muse Ember does not try to be the smartest model, but the cheapest one that is good enough — and on launch-day benchmarks it ties Claude Haiku 5.5 while costing about a third as much per task. That is a real shift for anyone running agents at scale. The honest test comes next: whether those numbers hold up on real workloads.
FAQ
How much does Muse Ember cost?
It costs $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004. Those rates apply up to 200,000 input tokens; above that, both triple.
Is Muse Ember better than Claude Haiku 5.5?
On the Artificial Analysis Intelligence Index, Ember scored 36 and Haiku 5.5 scored 36.2 — effectively a tie — while Ember costs about $0.14 per task against $0.41.
Can I run Muse Ember on my own hardware?
No. Meta keeps Ember's weights private, so it is only available through the Meta Model API, Muse Code and other hosted providers.
Sources
- https://research.meta.ai/blog/muse-ember?ref=digitaljetty.com
- https://talkwithtech.com/muse-ember-pricing-explained/
- https://www.eesel.ai/blog/meta-muse-ember
Related reading on Daily Tech & AI:
- Claude Fable 5.2 Is Here — Mythos Stays Locked
- Google TPU 8tx vs TPU 8i Max: Which One Wins?
- 5 AWS Launches That Make AI Agents Cheaper
Researched with AI assistance and reviewed for accuracy.
Follow on WhatsApp
Follow Daily Tech & AI on WhatsApp for the latest tech and AI updates: https://www.whatsapp.com/channel/0029Vb9XzdHDJ6H4TFNAl61r
Comments
Post a Comment