It is two in the morning, and a developer in Karachi is watching lines of code scroll past on her screen. She is not typing. An AI agent is doing the work for her — reading files, running tests, fixing small bugs, and calling one tool after another while she sips her chai and waits for the green checkmarks to appear.
Every one of those tiny jobs looks free. It is not. Each tool call, each file read, each small code fix is a separate request to an AI model, and the bill for a single night of agent work can quietly add up to more than the developer earns from the project itself.
On Monday, Meta decided to attack exactly that problem. The company unveiled Meta Muse Ember, the smallest and most efficient model in its Muse family, built for the repetitive work AI agents repeat all day — and priced so low that the maths of running agents suddenly changes.
What Is Meta Muse Ember?
Meta Muse Ember comes from Meta Superintelligence Labs, the team behind the company’s Muse models. Meta describes it as its smallest model that still delivers serious reasoning power, designed for what it calls “complex reasoning execution” — the unglamorous jobs that keep an agent moving.
It is also the first model trained using Muse Wonder Full, Meta’s frontier reasoning system. In plain terms, the big, expensive model acted as the teacher, and Ember is the student that learned to do the everyday work at a fraction of the cost.
The model understands text, images, audio and video, and it works with a context window of one million tokens. That is enough room to swallow an entire codebase or a thick pile of documents in a single request, then answer from all of it at once.
Meta Muse Ember Pricing: Built for the Bill, Not Just the Benchmark
The headline number is the price. Meta Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens, according to Meta. Cached input — the text the model has already seen — drops to just $0.004 per million tokens.
For short prompts, that puts Ember at the same price as OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5. The difference shows up when the work gets heavy: Meta says Ember delivers more intelligence per dollar than either rival in any reasoning setting.
Mark Zuckerberg put it simply when he announced the launch: “Releasing Muse Ember today — our smallest model that delivers higher intelligence per dollar than others in its class.” It is a sentence aimed squarely at developers who pay for every token their agents burn.
How Meta Muse Ember Compares to Claude Haiku 5.5 and GPT-6 Luna
Benchmarks are where Meta makes its boldest claim. On the Artificial Analysis Intelligence Index, version 5, Meta Muse Ember scores 36 points — the same as Claude Haiku 5.5 and comfortably ahead of GPT-6 Luna at 28, according to the figures Meta shared.
The cost gap is even wider. Meta puts Ember’s cost at about $0.14 per Artificial Analysis task, against $0.41 for Claude Haiku 5.5 and $0.15 for GPT-6 Luna, all running at maximum reasoning effort. In other words, similar brains, roughly a third of the bill.
| Benchmark | Meta Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| Intelligence Index v5 | 36 | 36 | 28 |
| Coding Agent Index | 46 | 37 | 31 |
| Terminal-Bench 4.0 | 40.9% | 39.2% | 16.4% |
| FrontierCode 2.0 | 27.2% | 34.9% | 25.2% |
| Cost per task | $0.14 | $0.41 | $0.15 |
Coding is where Ember looks strongest. It leads the Coding Agent Index with 46 points, ahead of Haiku’s 37 and Luna’s 31, and tops Terminal-Bench 4.0, an agentic coding test, with 40.9 percent.
Haiku still wins one round. It leads FrontierCode 2.0, which tests real-world software engineering, with 34.9 percent against Ember’s 27.2 percent — a reminder that no small model wins everywhere.
Ember also offers six reasoning levels — minimal, low, medium, high, xhigh and max — so developers can dial the thinking effort up or down depending on the job. A quick routing decision needs almost no reasoning; a tangled code fix gets the full treatment. Neowin has published the full benchmark breakdown for readers who want every number.
What Meta Muse Ember Is Actually Built For
Meta is not pitching Ember as the model that writes your novel or solves a research problem. It is the model for the hundred small jobs in between: calling tools in sequence, reading a file, running a test, spotting a bug, making one precise change, and moving on.
In early testing, Meta’s partners used it for everything from reviewing contracts to writing code. Long document parsing is another natural fit — the million-token window means an agent can hand Ember an entire policy manual or legal file and ask it to pull out exactly what matters.
That is also why the model matters beyond Meta. As AI agents spread from demos into real workplaces, the cost of each tiny step decides whether companies can afford to let agents run all day. A model that is three times cheaper for the same quality removes the biggest objection in the room.
Where You Can Use Meta Muse Ember Today
Meta Muse Ember is available now through the Meta Model API under the model ID muse-ember-1.0, and inside Muse Code under any pay-as-you-go or subscription plan.
It also runs on Google Cloud and Oracle Cloud, and it supports OpenAI’s Responses API specification, so developers can drop it into existing agent workflows without rebuilding them.
The timing is no accident. Ember arrived on the same day Anthropic launched its Claude Fable 5.2 and Mythos 5.2 models, and a day after Google unveiled its new TPU 8tx and 8i Max AI chips.
It has been a week in which every AI giant is fighting for the same developers, and the cheapest capable model may win more of them than the smartest one. Meta also announced the same day that its Muse agent is coming to Apple Watch and Wear OS, taking the assistant from the desk to the wrist.
Why Meta Muse Ember Matters — and What Happens Next
The AI race has spent two years chasing bigger, smarter, more expensive models. Ember points the other way. Meta’s bet is that the next wave of AI is not one giant brain answering one hard question, but thousands of small agents doing thousands of small jobs — and that wave needs a model priced like a utility, not a luxury.
The same week also brought NVIDIA’s GR00T 2, a robot model that learns by watching humans — another sign that agents, not chatbots, are where the industry is heading.
If Ember’s benchmarks hold up outside Meta’s own testing, expect rivals to answer fast. OpenAI and Anthropic both price their small models at the same $0.10 entry point, and neither will want to be known as the expensive option for agent work.
For now, the model to watch is not the one that tops the leaderboard. It is the one sitting on the efficiency frontier — no model scoring higher for less per task — quietly making it affordable to let an AI agent work through the night.
The developer in Karachi, watching those green checkmarks appear, may soon find the bill finally matches the magic.
