🔍 Intelligence Synthesis JP/EN
Ethics · Regulation · Technology — Pharma Practice Notes

AI Tech Flash ── 2026-11-11

Meta releases small AI model Muse Ember

Meta says it matches Haiku 5.5's index score at about 1/3 the cost

What changes

According to MarkTechPost, Meta has released Muse Ember as a small model for the repetitive steps inside agents. An agent is an AI that carries out several steps on its own.

In Meta's announcement, the model reaches the same index score as Claude Haiku 5.5 at about one third of the cost per task. For users running large volumes of work, that could change cost estimates. The figures come from Meta's own announcement.

Why it matters

VentureBeat reported that Muse Ember nearly matches Claude Haiku 5.5 at an estimated per-task cost of one third. According to MarkTechPost, Artificial Analysis summarized the result the same way.

In Meta's announcement, Muse Ember scored 36 on the Artificial Analysis Intelligence Index, level with Claude Haiku 5.5 at max effort. GPT-6 Luna scored 28. Artificial Analysis is an AI evaluation firm.

MarkTechPost also notes that Muse Ember trails Haiku 5.5 on a test of real code repositories. Its weights are private, and it can only be used on Meta's hosted services.

What it is

Muse Ember is a reasoning model released by Meta Superintelligence Labs. A reasoning model is an AI built to work through a problem step by step before answering. According to MarkTechPost, Meta calls it its smallest and most efficient model.

According to MarkTechPost, Meta lists tool calls, coding subtasks, routing, extraction and long document parsing as targets. The model accepts text, images, audio and video directly.

It is the fourth model under the Muse brand, after Spark, Glimmer and Wonder.

Who and where

Meta Superintelligence Labs built it. According to MarkTechPost, Muse Ember runs only on Meta's own hosted infrastructure. It is available through the Meta Model API and through Muse Code, Meta's coding agent.

The weights are private. Weights are the numerical data that make up a trained model. This means it cannot run on a local GPU or a laptop, and there is no quantized (compressed) version.

According to MarkTechPost, early-testing partners included Box, Cline, Oracle, Cloudflare and You.com. The outlet gives the release date as November 9, 2026.

Why it was built

According to MarkTechPost, Meta describes a division of labor among its models. Muse Wonder handles the hardest problems, and Muse Spark covers coding and agentic workflows. Muse Ember takes repeatable tasks at scale.

Meta's own framing is that Muse Wonder plans and Muse Ember handles the volume. The problem it targets is processing the small steps of agents cheaply and in large numbers.

How it works

According to MarkTechPost, Muse Ember is the first model Meta trained using Muse Wonder Full, its largest model. It can read screenshots, charts, scanned pages, voice memos and meeting recordings without a separate model. It also operates desktop apps from screenshots.

The context window is 1 million tokens. A context window is the amount of text a model can read at once, and a token is a unit of text processing. Output can reach 512,000 tokens, four times the 128,000 cap on both rivals.

Prices are $0.10 per million input tokens and $0.50 per million output tokens. According to MarkTechPost, Claude Haiku 5.5 and GPT-6 Luna list the same base rates. The outlet explains that the cost gap must come from the number of tokens used per task.

Key points

  1. In Meta's announcement, it scores 36 on the intelligence index, tying Claude Haiku 5.5, at $0.136 per task versus $0.411 for Haiku 5.5.
  2. It scored 37.8 on the agent-workflow test AutomationBench-AA. Haiku 5.5 scored 13.9 and GPT-6 Luna scored 20.2.
  3. Weights are private, and it is available only via the Meta Model API and Muse Code. It scored 27.2 on the real-code test FrontierCode 2.0-AA, below Haiku 5.5's 34.9.

Conclusion

In Meta's announcement, Muse Ember reaches Haiku 5.5-level index scores at a lower cost per task. It is strong on agent-workflow tests and weaker on real-code tests.

The figures come from Meta's announcement. According to MarkTechPost, some come from Artificial Analysis and some from Meta's internal tests.

Quotes and sources

“Meta’s proprietary Muse Ember nearly matches Claude Haiku 5.5 at 1/3 the estimated per-task cost”
── VentureBeat, 2026-11-10
“It scores 36 on the Artificial Analysis Intelligence Index, level with Claude Haiku 5.5 at max effort.”
── MarkTechPost, 2026-11-10
“Muse Ember is the first model Meta trained using Muse Wonder Full, its largest model.”
── MarkTechPost, 2026-11-10
“Weights are private, so there is no local GPU, laptop or quantized option.”
── MarkTechPost, 2026-11-10

Source articles

Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.