Meta launches Muse Ember, a cheap AI model built for repetitive agent tasks

What you could do with it

• A developer could run hundreds of repetitive Zapier-style business automations—like routing data, calling tools, and parsing long documents—at about a third of the cost of rival models ¹.
• A business could feed an entire codebase or document set into one prompt thanks to the model's 1 million token context window and get up to 512K tokens of output back ².
• A developer could build an agent that watches computer screens through screenshots and operates desktop apps without needing a separate vision model ².

What Muse Ember is for

Muse Ember is the fourth model under Meta's Muse brand, sitting below Wonder and Spark in capability. While the larger models handle planning and complex coding, Ember takes over the repeatable work that fills up an agent's schedule: tool calls, routing decisions, data extraction, and parsing long documents ¹².

It natively understands text, images, audio and video, so an agent can read screenshots, charts, scanned pages, voice memos and meeting recordings without switching to a separate model ². Meta also noted it can operate desktop apps by watching their screens .

Performance at a glance

On Meta's main intelligence ranking from Artificial Analysis, Ember scores 36 at maximum reasoning effort — matching Claude Haiku 5.5 and beating GPT-6 Luna, which sits at 28 ¹².

Its strongest area is automation workflows. On AutomationBench-AA, which tests 657 real Zapier business automations, Ember completes 37.8 percent of tasks, well ahead of Luna's 20.2 percent and Haiku's 13.9 percent ¹². On coding-specific benchmarks it is more mixed: it leads Terminal-Bench 4.0 at 40.9 but falls behind Haiku on FrontierCode 2.0-AA, scoring 27.2 to his 34.9 .

Price vs. rivals

Ember's standard pricing matches that of competing services: $0.10 for every million input tokens and $0.50 for every million output tokens. But Meta says each actual task costs about a third as much because Ember uses fewer tokens to get the same job done ². At maximum reasoning effort, Meta puts the cost at $0.136 per Artificial Analysis task, compared with $0.411 for Haiku and $0.152 for Luna ¹.

Long prompts widen the gap further. Haiku adds surcharges above 100,000 prompt tokens, while Luna doubles its input rate above 272,000 tokens. Meta lists no long-context tier for Ember ¹.

How to get it

Ember has closed weights, meaning you cannot run it locally on your own GPU or laptop. It is available only through the hosted Meta Model API and inside Muse Code, Meta's coding agent ¹³. It supports the OpenAI-compatible Responses API, so existing code that points at OpenAI should mostly work with a setting change . Early testing partners include Box, Cline, Oracle, Cloudflare and You.com . The model launched on November 9, 2026 .

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

Key facts

• Muse Ember scores 36 on the Artificial Analysis Intelligence Index, tying Claude Haiku 5.5 and beating GPT-6 Luna
• Costs about $0.136 per task, roughly one-third of Claude Haiku 5.5's $0.411
• 1 million-token context window with up to 512K output tokens — four times the rival cap
• Strongest on automation workflows (37.8% on AutomationBench-AA), weakest on mergeable real-code tasks (27.2% on FrontierCode)
• Available only via Meta Model API and Muse Code; weights are closed and no local option exists

Sources

1. Meta Releases Muse Ember, a 1M-Context Model for High-Volume Agent Work — marktechpost.com
2. Introducing Our Smallest, Most Efficient Model: Muse Ember
3. Meta Model API