Meta's smallest Muse model costs $0.10/$0.50 per million tokens, has a 1M-token context and ties Claude Haiku 5.5 on Artificial Analysis at about a third of the cost per task.
On November 9, 2026, Meta Superintelligence Labs launched Muse Ember ↗. Meta calls it its "smallest and most efficient model." It's built for "the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." TechCrunch ↗, The Verge ↗ and VentureBeat ↗ covered the launch the same day.
Ember is the fourth Muse model, after Spark (April), its open-weights twin Glimmer (August) and the frontier Wonder (October), TechCrunch notes ↗. It's also the first model Meta trained using Muse Wonder Full. Unlike Glimmer, Ember is proprietary: you reach it through an API, not a weights download.
Here's what's on the model page ↗:
Model id: muse-ember-1.0, available in the Meta Model API and Muse Code.
Context: a 1M-token window and up to 512K output tokens.
Inputs: text, images, audio, video and PDFs. Output is text only.
Endpoints: Responses, Chat Completions and Messages, per VentureBeat ↗.
Other routes: Google Cloud and Oracle Cloud, plus OpenRouter ↗ and Vercel AI Gateway ↗ on day one.
| Prompt size | Input / cached / output (per 1M tokens) |
|---|---|
| Up to 200K tokens | $0.10 / $0.004 / $0.50 |
| Over 200K tokens | $0.30 / $0.012 / $1.50 |
Prompts over 200K tokens cost three times as much on every line, so the 1M window isn't cheap if you fill it. For comparison, Meta lists Muse Spark 1.3 at $1.25 input and $4.25 output.
Meta's headline claim is that Ember delivers "more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting." That claim rests on the Artificial Analysis ↗ launch run. Ember scored 36 on the Intelligence Index at max reasoning, the same rounded score as Claude Haiku 5.5. It cost about $0.14 per task, against $0.41 for Haiku. Unrounded, it's 35.8 against 36.2, so Haiku is still slightly ahead. GPT-6 Luna scored 27.8 at about $0.15 per task.
Artificial Analysis explains where the savings come from: Ember has the same input and output prices as Haiku 5.5 and Luna, but its cached input "is 60% cheaper," and cached input makes up most of the cost in agent runs.
Selected scores at max reasoning, from Meta's post ↗:
| Benchmark | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| AutomationBench-AA | 37.8% | 13.9% | 20.2% |
| Terminal-Bench 4.0 | 40.9% | 39.2% | 16.4% |
| Coding Agent Index-AA | 46 | 37 | 31 |
| FrontierCode 2.0-AA | 27.2% | 34.9% | 25.2% |
| OSWorld 2.1 (partial credit) | 69.5% | 68.9% | 50.5% |
These are vendor-published numbers. Agents' Last Exam, OSWorld and PerceptionBench are Meta's own internal runs, and Meta's "Personal Superintelligence Index" is an internal composite. Each model ran in a different harness on Terminal-Bench: mini-swe-agent for Ember, Claude Code for Haiku and Codex for Luna. Ember loses on FrontierCode, which judges real engineering tasks on whether the code could be merged. That's the result to watch if you want it writing production code.
"Muse Ember offers nearly Muse Spark 1.3 (max) intelligence at a 17x cheaper price point," wrote Alexandr Wang ↗, Meta's Chief AI Officer.
"36 vs 37 on intelligence, $0.14 vs $2.39 per task... one point of smarts for seventeen times the price is a trade most people will take," replied Bruce Mi ↗.
"i don't need a genius to extract 50k invoice fields, i need something fast that doesn't burn my budget," said ucan ↗.
Others were skeptical. "Now retire the in-house 'Personal Superintelligence Index' and let the third-party scores carry it," posted AI Early Edition ↗.
And on openness: "why not just open source this then instead of open sourcing a multi generation old worse version of muse spark," asked pigeon ↗.
We couldn't pull Reddit threads for this story, so these reactions are from X only.
Ember is a subagent model, and you should judge it as one. Meta's own pitch is "Muse Wonder plans, Muse Ember handles the volume." If your agent sends hundreds of small calls per task, such as file reads, test runs, field extraction or routing, Ember's low cached-input price is where you'll actually save money. Matching Haiku on the Artificial Analysis index at a third of the cost per task is a real result, and it comes from an independent benchmark, not only from Meta's charts.
Don't swap it in as your main coder yet. Ember trails Haiku 5.5 by almost eight points on FrontierCode, and its Terminal-Bench lead was set in a different harness. For the main coding loop, the picks in our Claude Code alternatives guide and our ChatGPT vs Claude vs Gemini comparison still apply. Use Ember for the cheap work underneath.
Who should try it: teams already running multi-agent pipelines (our framework comparison covers the harness side), anyone parsing long PDFs or call recordings at scale, and OpenRouter users who can A/B it against Haiku or DeepSeek V4.1 Flash in an afternoon. If you need weights you can self-host, look at Meta's open Glimmer line or our open-source LLM ranking instead. For more on Meta's consumer side, see our report on Muse Wonder's active memory.