Meta's new model costs a third of Claude Haiku's price and still beats it on agent tasks
Muse Ember is Meta's smallest model yet, priced at 10 cents per million input tokens, and the company says it matches Claude Haiku 5.5's top score for about a third of the cost.
Meta released a new model called Muse Ember on November 9, and it isn’t pitching chatbot users at all. Muse Ember is built for the boring half of an AI agent’s job: calling tools, routing a request to the right place, pulling text out of a messy document, running the same coding subtask for the thousandth time. Meta’s own announcement calls it the smallest and cheapest model the company has shipped, and the pitch is blunt: a heavier model plans, Ember does the volume.
The price is the headline
Muse Ember runs $0.10 per million input tokens and $0.50 per million output tokens, with cached input priced at $0.004 per million. Meta says that works out to about $0.14 per task on the Artificial Analysis Intelligence Index, against roughly $0.41 for Claude Haiku 5.5 at its top reasoning setting. On Meta’s own chart, Ember matches Haiku’s best score at under a third of Haiku’s cost, and it beats every reasoning setting of GPT-6 Luna outright.
| Model | Cost per task (Meta’s figures) | Intelligence Index score |
|---|---|---|
| Muse Ember (max) | $0.14 | 36 |
| Claude Haiku 5.5 (max) | $0.41 | 36 |
| GPT-6 Luna (xhigh) | $0.15 | 27.8 |
I’d treat all of that as Meta’s homework, graded by Meta. These are Meta’s own benchmark runs against Artificial Analysis’s public leaderboard, not an independent test I’ve seen replicated elsewhere, and vendors have an obvious reason to pick the comparisons that flatter them. Meta also didn’t publish a parameter count, which it usually does for models it calls “smallest.” The announcement just says “smallest” and leaves it there.
What it’s actually built for
Agents spend most of their tokens on small, repetitive steps, not on one big thoughtful answer, and Ember is tuned for that pattern. It carries a 1-million-token context window and can output up to 512K tokens in one call, big enough to take in a full codebase or a stack of documents at once. Meta’s own numbers show it ahead of Haiku and Luna on AutomationBench, a business-workflow benchmark (37.8% versus 13.9% and 20.2%), and on Terminal-Bench 4.0 for agentic coding (40.9% versus 39.2% and 16.4%). It loses one comparison outright: Haiku scores higher on FrontierCode 2.0, a harder real-engineering benchmark, 34.9% to Ember’s 27.2%.
Ember also reads images, audio and video natively, so an agent can work from a screenshot, a scanned PDF or a voice memo without handing the job to a second model. Box, an early partner, said in Meta’s release that Ember could score a contract clause-by-clause against a negotiation playbook and land on the right risk rating. That’s a partner quote Meta picked for its own announcement, worth noting as such, not an independent review.
Where it fits in Meta’s lineup
Ember is the fourth model in Meta’s Muse line this year, after Spark in April, the open-weight Glimmer in August and the frontier Wonder model in October. It’s also the first model trained using the full version of Wonder, which Meta says is how it plans to keep producing smaller, cheaper models going forward. The pattern reads like OpenAI’s and Anthropic’s own mini-model pushes: once a lab has a frontier model, the next move is squeezing a cheap, fast version out of it for the agent work that doesn’t need the expensive one.
My read: the interesting number here isn’t the benchmark score, it’s $0.14 a task. If that cost holds up outside Meta’s own testing, it changes the math on running agents at real volume, the kind of workload where a few cents per call adds up fast across thousands of runs a day. I’d still wait for someone outside Meta to run the same comparisons before taking the “beats Haiku at a third of the price” claim at face value. Underdog made a similar case for small models being underrated when it shipped a 32MB voice model that beat a 41GB rival on accuracy; cheap and small keeps turning out to matter more than labs’ marketing used to suggest.
Muse Ember is live now on the Meta Model API and Muse Code as muse-ember-1.0.
Note on sources: I read Meta’s full announcement on research.meta.ai, including its benchmark tables and the Box partner quote, and TechCrunch’s report, which matches the primary page on pricing, dates and the Muse lineup. All benchmark numbers in this piece are Meta’s own; no outside lab has re-run them yet.