Meta released Muse Ember on November 9 — a model built not for the hardest problems, but for the repetitive sub-tasks AI agents execute hundreds of times per session: tool calls, routing, data extraction, coding fixes, and document parsing. On the Artificial Analysis Intelligence Index, it scores 36, identical to Anthropic's Claude Haiku 5.5, at roughly one-third the cost per task. On AutomationBench, which tests end-to-end business workflows, it scores 37.8% — nearly triple Haiku's 13.9%.
The price tag is the same — the bill is not
Here's the counterintuitive part: Muse Ember's list price is $0.10 per million input tokens and $0.50 per million output tokens — the same headline rate as Claude Haiku 5.5 and OpenAI's GPT-6 Luna. So where does the cost advantage come from?
Two places. First, cached input tokens — the bread and butter of agentic workflows where the same system prompt and conversation history get resent on every step — cost $0.004 per million tokens on Ember, versus $0.01 on Haiku 5.5. That's 60% cheaper on the tokens that dominate most agent bills.
Second, Ember is dramatically more concise. According to Artificial Analysis testing, Ember uses roughly 84,000 output tokens per Intelligence Index task. Haiku 5.5 uses around 226,000 — nearly three times as many. Since output tokens are billed at 5× the input rate, this alone slashes the per-task cost.
Where Ember wins — and where it doesn't
The benchmarks tell a clear story about what this model is for and what it isn't.
On agent-focused tasks, Ember dominates its weight class. Its 37.8% on AutomationBench — which tests full end-to-end business workflows like pulling records, calling tools in sequence, and updating systems across apps — is 2.7 times Haiku's score and nearly double Luna's 20.2%. On Terminal-Bench 4.0, which tests agentic terminal coding, Ember scores 40.9% against Haiku's 39.2% and Luna's 16.4%.
But Ember has real gaps. On FrontierCode 2.0, which measures whether a model's code changes would actually get merged, Haiku leads 34.9% to 27.2%. And on general knowledge, Ember is conspicuously weak: its AA-Omniscience accuracy is just 21.1%, less than half of Haiku's 34.7%. As FelloAI's analysis noted, Ember declines to answer 64% of knowledge questions — it prefers to say "I don't know" rather than guess. That's a feature for agent work where hallucinated tool calls are worse than no action at all, but it makes Ember a poor fit for anything that leans on recall.
Meta is explicit about the intended pairing: "Muse Wonder plans, Muse Ember handles the volume," the announcement states. Think of it as the workhorse behind a smarter orchestrator.
Why this matters for businesses running agents
For any company running AI agents at volume, the cost of small model calls is the single largest variable on the monthly bill. Not because individual calls are expensive — they're fractions of a cent — but because agents make dozens of calls per task, and context accumulates.
A detailed cost analysis from Crevio found that a single business agent doing 300 jobs per month costs roughly $74 in model tokens alone, with context re-transmission driving the bill far more than call count. An eight-step coding agent costs 6-7× more than a single-call equivalent due purely to context window growth at each step.
This is exactly the economics Ember is designed to disrupt. If you're running a customer support agent that handles 2,000 conversations a month, or a document processing pipeline ingesting a thousand invoices, the difference between $0.14 and $0.41 per task isn't academic — it's the difference between a model bill of $280 and one of $820. Scale that to five agents across different business functions and you're looking at thousands in monthly savings.
Meta also shipped Ember with a 1-million-token context window and 512K output tokens — enough to ingest a full codebase or document set in a single call. And it handles text, image, audio, video, and PDF inputs natively, meaning "agents can read screenshots, charts, and scanned documents, or work from voice memos, meeting recordings, and video without a separate model," as TechCrunch's Russell Brandom noted.
The bigger picture: a race to the bottom of the agent stack
Muse Ember slots into a pattern we've been tracking all year. When Meta launched Muse Wonder in October, the story was about competing at the frontier. Now the story is about competing at the floor — building the cheapest possible execution layer so that running agents at scale stops being a line item you worry about.
Every major lab is making the same move. OpenAI has GPT-6 Luna. Anthropic just dropped Haiku 5.5 to match the $0.10/$0.50 price point. Google has Gemini 3.8 Flash. The list prices are converging. What differs is efficiency — how many tokens the model burns to get the job done — and that's where Ember's architecture gives Meta an edge.
For business owners, this is unambiguously good news. The cost of the "worker" layer in AI agent architectures is falling faster than anyone expected. Companies like Uber learned the hard way that AI token costs can spiral without discipline. Models like Ember make that discipline easier to maintain — not because they're radically cheaper per token, but because they use fewer tokens to reach the same answer.
What to watch
The immediate question is whether Ember's agent benchmark scores hold up in production. AutomationBench numbers are impressive, but real business workflows are messier than benchmarks. Watch for early adopter reports in the coming weeks — particularly from companies already using Meta's Muse for Small Business platform, which now has a natural execution model to pair with Muse Wonder's planning capabilities.
The broader question is whether the "planner plus executor" architecture — a big model that thinks, paired with a small model that acts — becomes the standard pattern for business AI. Meta is betting it will. If Ember's per-task economics hold, that bet looks increasingly sound.
Sources
- Introducing Our Smallest, Most Efficient Model: Muse Ember — Meta Research
- Meta releases Muse Ember, a new lightweight model for agent execution — TechCrunch
- Meta's Muse Ember Matches Claude Haiku 5.5 On Artificial Analysis Intelligence Index At One-Third Of The Price — OfficeChai / Artificial Analysis
- Meta Muse Ember: Price, Benchmarks and the 200K Catch — FelloAI
- Meta's Muse Ember matches Claude Haiku at a third of the cost per task — The Next Web
- AI Agent Running Costs: What One Agent Really Costs per Month — Crevio
