Meta Launches Muse Ember: Its Smallest, Cheapest AI Model for High-Volume Agent Work
Key takeaway: Meta has launched Muse Ember, its smallest and cheapest AI model, at $0.10 per million input tokens and $0.50 per million output tokens. Built for agents’ repetitive work — tool calls, routing, coding subtasks, document parsing — it scores 36 on the Artificial Analysis Intelligence Index, level with Claude Haiku 5.5, at roughly a third of Haiku’s cost per task.
What Meta announced on November 9
Meta introduced Muse Ember on November 9, 2026, calling it the company’s “smallest and most efficient model.” It is available now on the Meta Model API and in Muse Code as muse-ember-1.0, and Meta says it is also on Google Cloud and Oracle Cloud.
Ember is designed for what Meta calls “complex reasoning execution” — the small jobs an agent repeats hundreds of times a day rather than the big one-off reasoning tasks: tool calls, coding subtasks, routing, extraction, and parsing long documents. It accepts text, images, audio, and video, with a 1-million-token context window and up to 512,000 output tokens.
What Muse Ember costs
The headline price is $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. One catch: requests over 200,000 input tokens are billed at triple rates — $0.30 input, $0.012 cached, $1.50 output — for the whole request, not just the part above the threshold.
On the Artificial Analysis Intelligence Index at maximum reasoning effort, Ember scores 36, versus 37 for its bigger sibling Muse Spark 1.3 — but the cost per index task is $0.14 for Ember against $2.39 for Spark. Ember also matches Claude Haiku 5.5 at 36, at about a third of Haiku’s $0.41 cost per task. The headline $0.10/$0.50 price matches Haiku and OpenAI’s GPT-6 Luna, so Ember competes on equal pricing with more efficient execution rather than undercutting the sticker price.
Muse Ember vs the bigger Muse models
| Model | Price per 1M tokens (in / out) | Intelligence Index | Cost per task | Best for |
|---|---|---|---|---|
| Muse Ember | $0.10 / $0.50 | 36 | $0.14 | Repetitive agent work: tool calls, routing, extraction |
| Muse Spark 1.3 | $1.25 / $4.25 | 37 | $2.39 | Harder reasoning, general agent work |
| Muse Wonder 1.0 | $4 / $20 | — | — | Frontier reasoning (Meta’s largest model) |
| Claude Haiku 5.5 | $0.10 / $0.50 | 36 | $0.41 | Fast, cheap general work |
Meta also notes Ember is the first model trained using Muse Wonder Full, its largest model — a teacher-student setup in which the flagship model helps train the small one. Meta has not disclosed a parameter count, and it has not released the model weights.
Why this matters
The agent economy has a cost problem, not a capability problem. A single agent run can make hundreds of model calls, and when every call goes to a flagship model, the bill grows faster than the value. That is why the industry is splitting into two tiers: big models for the hard thinking, small models for the repeated execution. Ember is Meta’s entry in that second tier — near-flagship scores on routine work at roughly one-seventeenth of Spark’s cost per task. The timing fits Meta’s recent Muse expansion into Spaces and Active Memory, which will drive far more agent calls than a chat window ever did.
What to do with it
If you build or run AI agents, Ember is worth testing on the jobs you currently route to a mid-size or flagship model: classifying and routing requests, extracting structured data from documents, running coding subtasks, and chaining tool calls. Start with tasks where a wrong answer is cheap to catch — a routing mistake you can log, an extraction you can validate — and measure the error rate against your current model before moving anything where mistakes are expensive.
Watch the 200,000-token threshold. Ember’s 1M-token context invites loading an entire codebase in one call, but a request that crosses 200K input tokens triples the price of the whole request. For very long inputs, splitting the work into smaller calls may cost less than one giant one.
Knowns, unknowns, and what to watch
Known: the price, the 1M-token context, the 512K-token output limit, availability on the Meta Model API, Muse Code, Google Cloud, and Oracle Cloud, and the benchmark scores Meta published with Artificial Analysis. Unknown: the parameter count, whether weights will ever be released, and how Ember performs outside the benchmarks Meta chose. What to watch: independent benchmark results, whether developers report the same cost savings in production, and whether rivals answer with cheaper execution-tier models.
Analysis: Ember’s real test is not the benchmark table — it is whether agents built on it stay reliable at scale. A model that scores 36 on an index but hallucinates a tool call in production is more expensive than a pricier model that gets it right. The next few weeks of developer reports will tell us if Meta’s bet pays off.
Frequently asked questions
What is Meta Muse Ember?
Muse Ember is Meta’s smallest and most efficient AI model, launched on November 9, 2026. It is built for the repetitive work AI agents do all day — tool calls, coding subtasks, routing, and document parsing — and it is available on the Meta Model API and in Muse Code as muse-ember-1.0.
How much does Muse Ember cost?
Muse Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Requests with more than 200,000 input tokens are billed at three times those rates for the whole request.
How does Muse Ember compare to Claude Haiku 5.5?
Both score 36 on the Artificial Analysis Intelligence Index at maximum reasoning effort, and both charge $0.10 per million input tokens and $0.50 per million output tokens. Ember’s cost per index task is $0.14, versus $0.41 for Haiku 5.5, according to Artificial Analysis data cited by Meta.
What is Muse Ember best used for?
Ember is best for high-volume, repetitive agent work: tool calls, routing between agents, information extraction, coding subtasks, and parsing long documents. It is not positioned as a replacement for Meta’s larger Muse Spark and Muse Wonder models on the hardest reasoning tasks.
Is Muse Ember open source?
No. Meta has not released Muse Ember’s weights, and it has not disclosed the model’s parameter count. It is available as a hosted API model on the Meta Model API, in Muse Code, and on Google Cloud and Oracle Cloud.
Sources: Meta AI Research — Introducing Our Smallest, Most Efficient Model: Muse Ember; Unite.AI — Meta Releases Muse Ember for Lower-Cost Agent and Coding Work; Zeniteq — Meta Launches Muse Ember for High-Volume Agent Work; Eesel AI — Meta Muse Ember explained: specs, uses, and how to call it. Related reading on USABite: AI Agents for Business Automation and AI-Powered Digital Products Explained.
