Fourteen cents per task versus forty-one. That is the gap between Muse Ember, the small model Meta unveiled on 9 November 2026, and Claude Haiku 5.5 on Artificial Analysis's Intelligence Index, where both score 36 at their top reasoning setting. At the same per-token price, Ember uses roughly a third of the output tokens.
This is not a chat model. Meta built it for the repetitive work agents actually do: tool calls, coding subtasks, request routing, data extraction and long-document parsing. In other words, the work that inflates the bill in any multi-agent system.
What Muse Ember brings
It is the smallest model in the Muse family and the first trained using Muse Wonder Full. It takes text, images, audio, video and PDF, has a one-million-token context window and can produce up to 512,000 output tokens. It has six reasoning levels, from minimal to max.
Pricing has two tiers:
| Prompt | Input | Cached input | Output |
|---|---|---|---|
| Up to 200,000 tokens | $0.10 | $0.004 | $0.50 |
| Over 200,000 tokens | $0.30 | $0.012 | $1.50 |
Prices per million tokens.
It is available through the Meta Model API and Muse Code, and Meta also lists it on Google Cloud and Oracle Cloud. Meta has said nothing about releasing the weights.
How to try it
The Meta Model API follows the Responses API format, so all you need is a key from dev.meta.ai and the model id muse-ember-1.0:
curl https://api.meta.ai/v1/responses \
-H "Authorization: Bearer $MODEL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "muse-ember-1.0",
"input": "Classify this ticket as billing, outage or feature request: ..."
}'Meta's quickstart also uses the OpenAI Python SDK pointed at https://api.meta.ai/v1, which makes it easy to slot into an existing model router.
Against Claude Haiku 5.5 and GPT-6 Luna
Its two natural rivals charge the same per token for short prompts ($0.10 input and $0.50 output per million), so the difference comes down to how many tokens each one burns and what it gets right. These are Artificial Analysis and Meta figures, all at maximum reasoning:
| Muse Ember | Claude Haiku 5.5 | GPT-6 Luna | |
|---|---|---|---|
| Intelligence Index (AA) | 36 | 36 | 28 |
| Cost per index task | $0.14 | $0.41 | — |
| Coding Agent Index (AA) | 46 | 37 | 31 |
| FrontierCode 2.0 (AA) | 27.2% | 34.9% | 25.2% |
| Terminal-Bench 4.0 (AA) | 40.9% | 32.8% | — |
| AA-Omniscience accuracy | 21.1% | 34.7% | — |
In the Coding Agent Index, Ember runs inside Muse Code, Haiku inside Claude Code and Luna inside Codex: it compares whole agents, not just models.
Where it wins and where it loses
It wins on cost per agentic task. Matching Haiku 5.5's score for a third of the money is exactly what matters in an orchestrator that fans out dozens of subagents per request. It also leads on Terminal-Bench 4.0, which measures real work in a terminal, and its cached input is 60% cheaper than its rivals', which adds up for agents that resend the same context every turn.
It loses on hard coding. On FrontierCode 2.0, the real-world programming test, Haiku 5.5 is almost eight points ahead. If the subagent writes code that ends up in production, Ember is not the obvious pick.
It knows little, and says so. On AA-Omniscience, which measures factual knowledge, it gets 21.1% of questions right against Haiku 5.5's 34.7%, but it declines to answer 64% of them and its hallucination rate is about half of Haiku's. For extraction and routing with the data in the prompt, that is a feature; for questions asked without context, it is a problem.
Bear in mind that several benchmarks in Meta's own table are internal (Agents' Last Exam, Personal Superintelligence Index), and partner figures from You.com or Cloudflare are published by Meta without independent verification. That is why the table above sticks to Artificial Analysis measurements.
Who it is for, and who it is not for
Yes, if you run an agent system with heavy volumes of tool calls, classification, extraction or long-document reading, and you are paying for it with a small model. It is the clearest candidate for an A/B test in your router: same price per token, fewer tokens and a better agentic score.
No, as a drop-in replacement for the model that writes your code, where Haiku 5.5 does better, nor for answers that rely on the model's own knowledge without supplied context. And if you need open weights or self-hosting, nothing has been announced.
Sources
- Muse Ember, Meta, 9 November 2026.
- Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index, Artificial Analysis, 9 November 2026.
- Muse Ember, Simon Willison, 10 November 2026.
- Quickstart, Meta Model API documentation.