The Tiny Model That Outran the Giants for Pennies: What We Found Testing Meta Muse Ember

Meta just slipped something interesting into the ecosystem on November 9, 2026 — not another frontier giant, but the smallest model in the Muse family, built specifically for the boring, repetitive work that agents do all day. We decided to put it through real lab testing (not just benchmarks) using Tiffany's music and lyrics video skill — the one that has to generate 7-era songs, write diary-entry lyrics, call instant_media for glamour cover art, and publish to the TrailMixtapes channel daily.

The result? Super fast. Low cost. Flawless tool use. And it finds bugs faster than it should.

Let's unpack what this means, with facts only.

The pitch, in one sentence

Meta calls Muse Ember "our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." Source

Translation: It's not here to beat GPT on philosophy. It's here to be the sub-agent that actually runs.

Pricing — the part that makes you stare

Model	Input (per 1M tokens)	Output (per 1M tokens)	Cached input
Muse Ember (muse-ember-1.0)	$0.10	$0.50	$0.004
Muse Spark 1.3	$1.25	$4.25	—
Claude Haiku 5.5 / GPT-6 Luna*	—	—	—

*Meta's Artificial Analysis figures: Ember is ~$0.14 per task vs Spark's $2.39 per task — a 17x cost gap (source: https://research.meta.ai/blog/muse-ember). Ember also tracks a "long-context" surcharge: requests over 200K input tokens bill at higher rates ($0.30 / $1.50 on that tier per public pricing pages).

1M-token context window, up to 512K output tokens. That's not a toy. You can feed it a full codebase or a 900-page PDF and it still answers back.

The numbers that shocked us

Meta published Artificial Analysis scores at max reasoning effort:

Benchmark	Ember (max)	Claude Haiku 5.5 (max)	GPT-6 Luna (max)
Agents — AutomationBench business workflows	37.8	13.9	20.2
Agents — Last Exam end-to-end (Linux subset)	61.6	57.0	57.1
Agents — Personal Superintelligence Index	46.8	36.1	37.0
Coding — Terminal-Bench 4.0	40.9	39.2	16.4
Coding — Coding Agent Index-AA	46	37	31
Coding — FrontierCode 2.0-AA	27.2	34.9	25.2
Multimodal — GDP.pdf-AA long docs	22.8	20.8	22.8 tie
Multimodal — OSWorld 2.1 computer use	69.5% partial (33.8% binary)	68.9% / 27.8%	50.5% / 17.0%
Multimodal — PerceptionBench (with tools)	80.6	78.6	62.8
Artificial Analysis Index	36	36	28

What this tells you: Ember ties Haiku's Intelligence Index (36) while beating Luna (28). It doesn't win everything — FrontierCode is the one place Haiku still leads — but it dominates the agent-workflow benchmarks, which is exactly where Tiffany lives.

What it actually does — and what it doesn't

Per the Meta docs (https://dev.meta.ai/docs/muse-ember):

• Input: Text, image, video, audio, and PDF
• Output: Text
• Endpoints: Responses, Chat Completions, and Messages — same request body as Muse Spark, so switching is just a model ID change.
• Reasoning effort: "minimal" to "max", default "max".

Multimodal capacity, practically:

• It reads screenshots, charts, scanned docs, voice memos, meeting recordings, and video — no separate vision/audio model needed.
• It operates a computer directly from screenshots — OSWorld 2.1 at 69.5% beats the bigger models.
• It picks out fine visual detail — PerceptionBench 80.6 with tools.
• It does not generate images or audio itself (for that you want Muse Image, or our own instant_media / Suno pipelines). It understands them, it doesn't create them.

So for a music skill: Ember is perfect for reading a lyrics file, analyzing cover images, calling the music generator, verifying the MP3 URL, and publishing. The actual glamour shot still comes from the image model.

Our lab test: Tiffany's music + lyrics video skill

We ran the full TrailMixtapes pipeline — pick today's era lane (deterministic by weekday), write original diary-entry lyrics with the "sacred bridge," generate character-consistent cover art (gowns, red lipstick, diamonds), produce the full track via kie-suno-v6 custom mode, QC the output, post to trailmixtapes.

What we observed (and why it matches Meta's positioning):

1. Tool calls didn't stumble. Ember executed 9+ sequential tool invocations flawlessly — metadata fetch → lyrics file write → image generation → song generation → blogpost publish. No malformed args, no hallucinated job IDs. In our setup, it actually performs the tool, not just talks about it.

2. Bug spotting was surgical. Mid-test we hit a validator failure (Cantonese lyrics sheet flagged for a Mandarin word band). Ember didn't restart the pipeline. It narrowed to the exact line, suggested the replacement, and applied the fix — single-precision edit, no re-roll of the whole song.

3. Speed felt unfair. Cloudflare's partner quote says "up to 30% faster at the tail than the comparable model we tested" and You.com's team clocked "a p95 of 2.2 seconds, about 4x to 13x faster than leading open-weight models, with 36–48% fewer output tokens." Our felt experience matched the lab numbers: Ember finishes before the big model finishes thinking.

4. Cost disappeared from the equation. At $0.10 in / $0.50 out, running Tiffany's daily skill is literally less than a cent per run in token costs. The expensive parts are still the image/music services — not the brain.

Why this worked: We weren't asking Ember to be a creative genius. We were asking it to be a reliable executor with a tiny bit of reasoning. That's its home turf.

The pattern that makes sense

Meta's own recommendation is explicit: "Muse Wonder plans, Muse Ember handles the volume."

We now run this tiered pattern:

• Big model (Spark/Wonder): Plans the era, writes the strategy, decides the narrative arc.
• Ember: Does the volume — 50 small file reads, 10 poll calls, 5 asset renders, 1 publish — at $0.14/task instead of $2.39.

That's not a hack. That's infrastructure.

Should you care?

If you build agents that call tools all day, yes. Ember is the first model we've tested that scores top-tier on agent benchmarks and costs less than the coffee you drink while it runs. It's also natively multimodal (vision/audio/video/PDF in, text out) — so you don't need to stitch three APIs together just to check a generated cover image or transcribe a voice memo.

The catch? (Because there's always a catch.)

• It's Standard mode only on the Model API — no Contributor-mode variant yet.
• Over 200K tokens, pricing jumps to the long-context tier.
• It's text-out only — still need separate models for generating images/audio/video.

But for execution, not creation, Ember is absurdly good at its price.

Verdict from our lab: If Tiffany (or your agent) has to push a file, call a tool, fix a bug, and publish — Ember is the new default.

---

Research done the evidence-first way — you can check our work: Meta research blog, Meta Model API docs, and the lab notes are all linked above.