Meta launched Muse Ember on Monday, its smallest and most efficient model yet, built for the repetitive work AI agents do all day. The model matches Claude Haiku 5.5’s best score on Artificial Analysis’s independent index at about a third of the cost per task, and beats OpenAI’s GPT-6 Luna at every reasoning setting Meta measured. It costs $0.10 per million input tokens and $0.50 per million output tokens, and it is available now on the Meta Model API and in Muse Code.
Built for the work agents repeat
Agents spend most of their tokens on many small, structured steps, and Ember is built for exactly that: tool calls, coding subtasks, routing requests, extracting data from long documents, and parsing. With a 1 million-token context window and up to 512,000 output tokens, it can take in a full codebase or document set in a single call.
It also reads text, images, audio and video natively, so agents can work from screenshots, charts, scanned documents, voice memos, meeting recordings and video without a separate model. Meta’s early partner Box put it to work scoring contracts against a negotiation playbook clause by clause, and screening candidates against a detailed hiring brief.
The price, and what it buys
Ember lists at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. That puts it far below the frontier models it is measured against: Anthropic’s Claude Fable 5.2, launched the same day, costs $8 and $40 per million tokens, eighty times Ember’s input price.
The comparison Meta wants is cost per task, not cost per token. On Artificial Analysis’s index, Ember scores 36 at $0.14 per task, against 37 for its own Muse Spark 1.3 at $2.39 per task — nearly the same score at about a seventeenth of the cost. No model on the index scores higher for less per task, Meta says.
The benchmark table, including one Haiku wins
Meta’s own table has Ember ahead of Haiku 5.5 and GPT-6 Luna on most tests it published. On AutomationBench, the end-to-end business workflow test, Ember scores 37.8 against 13.9 for Haiku and 20.2 for Luna. On Terminal-Bench 4.0, the agentic coding test, it scores 40.9 against 39.2 and 16.4.
The table is not a clean sweep, and Meta publishes the row it loses. On FrontierCode 2.0, the real engineering test, Haiku scores 34.9 against 27.2 for Ember and 25.2 for Luna. On the overall Artificial Analysis index, Ember ties Haiku at 36, with Luna at 28.
Trained by Wonder, paired with Wonder
Ember is the first model trained using Muse Wonder Full, Meta’s frontier model, which it introduced in October. When Meta launched Wonder, it said it would use its biggest models to build smaller ones for everyone, and Ember is the first result. For the hardest problems, Meta says to pair the two: Wonder plans, Ember handles the volume.
The pairing is a pricing strategy as much as an engineering one. A customer-support agent could have Wonder set the rules for handling an issue, then send Ember to apply those rules across tens of thousands of cases at $0.14 a task. That is the use case Meta is selling: frontier thinking once, cheap execution everywhere.
Safety checks before release
Before release, Meta evaluated Ember under its Superintelligence Scaling Framework, covering chemical and biological, cybersecurity and loss-of-control risks. Across the categories assessed, the model stayed below the framework’s high-risk thresholds, and it received the same core safety training and testing as Meta’s larger models.
Meta also tested how Ember behaves when acting on a user’s behalf: whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages and documents that try to redirect it, and seeks human input before high-impact actions.
Our Read
Meta launched its cheapest model on the same day Anthropic launched its most capable one, and the two point in opposite directions. Anthropic cut its flagship price by a fifth to sit below OpenAI’s GPT-6 Astra at $10 and $50, while Meta is betting the volume business belongs to models that cost almost nothing per task.
Google, meanwhile, is moving the other way: its Gemini 4 Argon API price doubles on December 1. The three labs are now charging in visibly different directions, and the cheapest capable model on the market is suddenly the one built for the work agents repeat all day.
The honest caveat is that Meta’s table is vendor-reported, and the one row that matters most for coding — FrontierCode 2.0 — goes to Haiku. Ember wins on price and on agent workflows, not on hard engineering. Whether developers accept that trade-off will show in usage, not in launch-day charts — and Meta says more efficient models are in training.
Why It Matters
Ember changes what it costs to run agents at scale. At $0.14 per task, a company can afford to put an agent on work that was never worth automating before — reviewing every contract, screening every candidate, routing every request — and Meta’s bet is that this volume, not frontier capability, is where the next wave of AI spending goes.
For rivals, the pressure is on price, not capability. Anthropic already answered with a cheaper flagship the same day, and Google is raising Argon instead. If Ember’s cost advantage holds up in independent testing, the cheapest capable model becomes the default for agent work — and the frontier labs start competing on cents per task.
Background
Ember is the fourth model in Meta’s Muse line, after the mid-weight Spark in April, its open-source twin Glimmer in August, and the Wonder frontier model in October. Meta’s consumer Muse Agent launched in September, and Anthropic filed to go public on October 26, with its roadshow able to start on November 10.
Source: Meta’s announcement and TechCrunch’s report
0 Comments