Meta has launched Muse Ember, its smallest and cheapest AI model yet, aimed at the lightweight models developers already use for agent work. According to Meta, Ember matches Anthropic’s Claude Haiku 5.5 on its best benchmark score at about a third of the cost per task, and beats OpenAI’s GPT-6 Luna at every reasoning setting. Announced on Monday 9 November by Meta Superintelligence Labs, it is available now through the Meta Model API and in Muse Code.
Ember is not built to be the smartest model in Meta’s lineup. It is the one agents call hundreds of times a day, designed for “complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.”
The cost comparison at the centre of the launch
The headline numbers come from Artificial Analysis, an independent benchmarking firm. On its Intelligence Index, Muse Ember scores 36 — the same as Claude Haiku 5.5 at its best setting, and well ahead of GPT-6 Luna’s 28. The difference is what each completed task costs: about $0.14 per task for Ember, against roughly $0.41 for Haiku and $0.15 for Luna, on Artificial Analysis estimates at list pricing.
That gap comes from the price list: $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Meta also notes Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task. The caveat: the cost figures are calculated at list prices, and the chart is Meta’s own presentation of Artificial Analysis data — though the index score is independently measured and confirmed on Artificial Analysis’ own model page.
Where it wins — and where it does not
Ember’s strongest published result is AutomationBench, an end-to-end test of business workflows: 37.8 against 13.9 for Haiku and 20.2 for Luna. It also leads on Agents’ Last Exam (61.6 to 57.0 and 57.1), on Terminal-Bench 4.0 agentic coding (40.9 to 39.2 and 16.4), and on the Coding Agent Index (46 to 37 and 31).
It does not win everywhere. On FrontierCode 2.0, a test of real engineering tasks judged on whether the work could actually be merged, Haiku leads at 34.9 against Ember’s 27.2, with Luna at 25.2. On long-document work measured by GDP.pdf, Ember ties Luna at 22.8, ahead of Haiku’s 20.8. The pattern fits the design: built for volume, not the hardest single problem. Meta’s own advice is to pair it with Muse Wonder: “Muse Wonder plans, Muse Ember handles the volume.”
What it can actually do
Ember takes text, images, audio and video as input, so one agent can read screenshots, charts, scanned documents, voice memos or video without a second model. Its context window holds one million tokens, with up to 512,000 output tokens, and it supports Meta’s Responses API, built for multi-step agent workflows.
Early testers put it to work on the jobs Meta is targeting. Box used it to score contracts against a company’s negotiation playbook clause by clause. Cloudflare reports it ran up to 30% faster on its slowest responses and handled investigations spanning several code repositories. You.com measured a p95 latency of 2.2 seconds and says Ember used 36–48% fewer output tokens than the leading open-weights models it tested. These are Meta-selected testimonials, not independent audits.
The Muse family, rounded out
Ember is the fourth Muse model, following the mid-weight Spark (April), its open-source twin Glimmer (August), and the frontier Wonder model (October). Meta says Wonder is for the hardest problems, Spark for coding and agentic workflows, and Ember for repeatable coding and agentic tasks at scale. By comparison, Muse Spark 1.3 scores one point higher on the index, at 37 — but costs about $2.39 per task, roughly seventeen times Ember’s $0.14. Ember is also the first model Meta says was trained using Muse Wonder Full, its most powerful system. Meta has not disclosed its parameter count, and unlike Glimmer, its weights are not being released: it is a proprietary, API-only model.
Why this matters
Ember arrives in a week crowded with AI launches — Thinking Machines’ Inkling 2 Small and OpenAI’s GPT-6 Cyber landed days earlier — but its argument is different. The others compete on capability; Ember competes on the bill. If agents run millions of small jobs a day, the cost of each job matters more than the brilliance of any single one, and Meta is betting that a model good enough at a tenth of the price wins the volume.
“Our goal is personal superintelligence,” Meta wrote in its announcement. “Bringing it to everyone takes more than frontier capability. It takes frontier capability at a price that scales.”
