Editorial illustration for Meta Releases Muse Ember, Lightweight Agent Model at 10 Cents Per Million Tokens
Meta Muse Ember: 10-Cent AI Model Beats Claude, GPT
Meta Releases Muse Ember, Lightweight Agent Model at 10 Cents Per Million Tokens
Meta Superintelligence Lab put out a new model Monday called Muse Ember, and the pitch is almost entirely about price. At 10 cents per million input tokens, Ember undercuts Anthropic's Haiku and OpenAI's Luna by a wide margin, while still landing close to open-source rivals like DeepSeek V4.1 Flash and GLM 5.3 Flash on cost. Meta is leaning on Artificial Analysis's composite benchmark score to argue Ember beats those frontier models on both price and intelligence, a combination that's rare in model releases this size.
The design logic is straightforward: Ember isn't meant to work alone. Meta built it as the workhorse half of a two-model setup, paired with a larger system like Muse Wonder, which handles planning while Ember grinds through repetitive execution, tool calls, coding subtasks, routing, extraction, long document parsing. Meta also built in multimodal handling, so Ember isn't limited to text input. That matters for the kind of high-volume agent work Meta says it's targeting, where a single heavier model sets the rules and a cheaper model applies them across thousands of cases.
Meta Superintelligence Lab released a new AI model on Monday called Muse Ember, the lab’s lightest and most cost-efficient AI model yet. In its release, Meta said Ember was “designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.”
Why this matters Meta's pitch here isn't about raw capability, it's about unit economics. Ten cents per million input tokens is cheap enough to change how teams architect agent pipelines, not just which model they pick. If Ember really matches or beats Haiku and DeepSeek V4.1 Flash on Artificial Analysis's composite score at that price, it becomes the default choice for the unglamorous work that eats most agent budgets: routing calls, parsing documents, chewing through coding subtasks nobody wants to pay frontier rates for. That's worth watching closely for founders building anything that fires off thousands of small model calls a day, since the margin difference between 10 cents and a frontier price tag compounds fast at scale.
Still, benchmark comparisons chosen by the vendor releasing the model deserve a raised eyebrow. Meta picked the metric, Meta picked the competitors, and "cost-competitive with open-source" is a lower bar than "better than." We'd want independent numbers on latency and real task accuracy before treating Ember as a default pick, not just a cheap one.
Related Reading
Browse all Meta news →
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic Meta’s Muse Tracks Your Email and Bank Data While You Work
AI Tools & Apps
Related Topic Meta Faces Lawsuit Over AI Training Data, Facial Recognition
Policy & Regulation
From Archives Baidu ERNIE Evolve Self-Reports 33.7% Throughput Gain Over NanoChat, NanoGPT
From Archives NVIDIA Isaac Lab Lets Robots Rehearse Delicate Work in Simulation at CoRL
From Archives Brain-as-computer metaphor shaped AI but may mislead consciousness debate
From Archives Meta’s Muse AI creates custom app spaces by letting users describe them
Open Source
Common Questions Answered
What is the pricing advantage of Meta's Muse Ember compared to competing models?
Muse Ember is priced at 10 cents per million input tokens, which significantly undercuts Anthropic's Haiku and OpenAI's Luna while remaining competitive with open-source alternatives like DeepSeek V4.1 Flash and GLM 5.3 Flash. This aggressive pricing strategy positions Ember as a cost-effective option for teams looking to optimize their AI infrastructure expenses without sacrificing performance.
What specific tasks is Muse Ember designed to handle according to Meta?
Meta designed Muse Ember specifically for complex reasoning execution and repetitive agent tasks including tool calls, coding subtasks, routing, extraction, and long document parsing. These capabilities make it particularly suited for the unglamorous but computationally intensive work that typically consumes the majority of agent budgets in production systems.
How does Muse Ember perform on benchmark scores compared to Haiku and DeepSeek V4.1 Flash?
According to Artificial Analysis's composite benchmark score, Meta claims that Muse Ember matches or beats both Anthropic's Haiku and DeepSeek V4.1 Flash on intelligence metrics. This performance parity combined with Ember's significantly lower cost creates a rare market advantage where the model achieves both superior price and comparable intelligence.
Why does Muse Ember's pricing model matter for agent pipeline architecture?
At 10 cents per million input tokens, Muse Ember's pricing is cheap enough to fundamentally change how teams design their agent pipelines rather than just influencing which model they select. This cost efficiency makes it economically viable to use Ember as the default choice for high-volume, repetitive tasks that previously represented significant operational expenses.
Further Reading
- Meta releases Muse Ember, a new lightweight model for agent execution - TechCrunch
- Meta launches Muse Ember to take on Claude Haiku and GPT-6 Luna - The New Stack
- Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index - Artificial Analysis
- Introducing Our Smallest, Most Efficient Model: Muse Ember - Meta AI Research
- Muse Ember Evaluation Report - Meta AI Research
