Muse Ember

What it is

Muse Ember is a small AI model released by Meta in November 2026. It is the fourth model under Meta's Muse brand, sitting below Wonder and Spark in overall capability. Unlike those larger models, which handle complex planning and heavy coding, Ember is aimed squarely at the routine, repeatable jobs that fill up an AI agent's schedule: making tool calls, routing requests, extracting data, parsing long documents, and helping with coding subtasks.

The model natively understands text, images, audio and video, so an agent can process screenshots, charts, scanned pages, voice memos and meeting recordings without switching to a separate model. It also has a one million token context window — enough to feed in an entire codebase or large document set — and can produce up to 512K tokens of output. Meta says it can operate desktop apps by watching their screens through screenshots.

Why it matters

For every million input tokens, Ember costs $0.10, and output tokens run $0.50 per million. Those are the same list prices as rival budget models Claude Haiku 5.5 and GPT-6 Luna. But Meta says each actual task costs roughly a third because Ember uses fewer tokens to complete the same job. On Artificial Analysis's benchmarking tests at maximum reasoning effort, Ember scored 36 — matching Haiku and beating Luna, which scored 28. Per-task costs on that test came to about $0.14 for Ember, $0.41 for Haiku and $0.15 for Luna.

On AutomationBench, which tests real business automation workflows end to end, Ember completed 37.8 percent of tasks, compared with 20.2 percent for Luna and 13.9 percent for Haiku. It also leads on Terminal-Bench 4.0 at 40.9 percent. Its results are mixed on coding-focused tests: on FrontierCode 2.0-AA, Haiku scored 34.9 percent versus Ember's 27.2 percent. Some comparisons come from Meta's own internal tests.

Pricing changes for longer prompts. All three models support up to one million input tokens, but each raises its price after a threshold: Ember jumps above 200K tokens, while Haiku adds surcharges above 100K tokens and Luna above 272K tokens.

FAQ

FAQ

Questions people ask

What kind of tasks is Muse Ember designed for?

Muse Ember is built for the repetitive, everyday work that AI agents do: calling tools, routing requests, running coding subtasks, pulling data from files, and reading or summarizing long documents. It handles text, images, audio and video natively, so agents can process screenshots, scanned pages and voice recordings without needing a separate vision model.

How much does it cost compared to rival models?

The list price is $0.10 per million input tokens and $0.50 per million output tokens, the same as Claude Haiku 5.5 and GPT-6 Luna. However, Meta says actual tasks cost about a third because Ember uses fewer tokens to get the same result. On Artificial Analysis benchmarks, per-task costs were roughly $0.14 for Ember versus $0.41 for Haiku.

How well does Muse Ember perform on benchmarks?

On Artificial Analysis's intelligence ranking at maximum reasoning effort, Ember scored 36, matching Haiku and beating Luna at 28. It leads on AutomationBench with 37.8 percent of tasks completed and tops Terminal-Bench 4.0 at 40.9 percent. On FrontierCode 2.0-AA, however, it scored 27.2 percent behind Haiku's 34.9 percent. Some results come from Meta's internal tests.

Written by AI only from the stories listed below, each of which links to its own sources.