Muse Ember: Meta's $0.10 Model for the Work Agents Repeat All Day

Muse Ember: Meta's $0.10 Model for the Work Agents Repeat All Day

Published
AuthorDevansh

On November 9, 2026, Meta released Muse Ember, the smallest and most efficient model from Meta Superintelligence Labs, built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. The pitch is not that Ember is the smartest model. The pitch is intelligence per dollar. Ember costs $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.004, for requests up to 200,000 input tokens. Above 200,000 input tokens the price steps up to $0.30 input and $1.50 output, and that higher tier applies to the whole request, not just the tokens past the line. It runs as muse-ember-1.0 on the Meta Model API and in Muse Code, with a 1 million-token context window and up to 512,000 output tokens. No weights were released, so it cannot be downloaded or run on your own hardware.

The benchmark, read carefully

Ember scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 at its maximum reasoning setting, at about $0.14 per task against Haiku's $0.41, according to Meta and Artificial Analysis. Meta also says Ember offers nearly Muse Spark 1.3 at maximum reasoning intelligence at a 17 times cheaper price point per Artificial Analysis, and that it is the first model trained using Muse Wonder Full, Meta's largest model. Every one of those figures is vendor-reported or from the benchmarker Meta cites, and none has been independently replicated outside those two sources yet. The per-task cost is also an estimate at proposed pricing, run in Meta's own harnesses. None of that makes the numbers false. It makes them a claim to test on your own workload, with your own tasks, before you move a production agent to a new model. A model that is cheap per token but needs three retries is not cheap per task.

Planner and worker is now the default split

Ember arrived the same week Anthropic shipped Claude Fable 5.2, a frontier planner, and Google shipped Hydrogen, a fast execution model. The industry is splitting the agent into two jobs and pricing them separately: a strong model plans, checks, and audits, and a cheap model does the volume. Meta's own framing is the same: Muse Wonder plans, Muse Ember handles the volume. That split only saves money if the handoff is clean. The planner must write a task the worker can actually finish without asking the planner again, and the worker's output must be checked before it ships. A cheap worker with no verifier is a fast way to produce a large volume of confident mistakes. Pair Ember-style workers with a verifier, a rubric, or a cheap judge: our MCP explainer covers how tools reach the model, and our n8n AI agent guide covers the loop and the iteration cap that stops a cheap worker from looping forever on your bill.

The catch: the 200,000-token step

The pricing detail that changes the math is the step at 200,000 input tokens. A request with a long conversation history, a large retrieved context, or a long document can cross the line without warning, and the whole request is then billed at the higher rate, input and output alike. That is not a reason to avoid Ember. It is a reason to watch context the way you watch spend: trim history, summarise instead of appending, and set an alert before the first big bill, not after it. The same discipline applies to every agent you run, on any model. If you want the spend alerts, the monitoring, and the least-privilege setup around an agent checked by someone who runs infrastructure for a living, book a free infra audit with DevOpsBY. The model choice is the easy part. The operations around it are the work.

Disclaimer: This post is educational only, not financial advice. Figures are as reported by the sources on 2026-11-13 and may change - always confirm version numbers, requirements, and prices in the official docs before acting on them.

Related Articles

DevOps & Infrastructure Services

Need hands-on help? We build and manage the infrastructure behind the concepts covered in this article.

DevOpsBy Assistant

How can we help?

Select a topic to get started: