Meta releases Muse Ember to do agent grunt work for cheap

If you've ever built an AI agent, you know the painful part isn't generating the master plan. It's the mountain of grunt work that comes after.

Running terminal commands, checking diffs, and parsing thousands of images can burn through an API budget in minutes. Meta clearly did the math, releasing what it calls its smallest and most efficient multimodal model yet: Muse Ember.

Ember isn't trying to be an all-knowing genius. Instead, Meta designed it strictly as a high-volume workhorse built to tackle the repetitive, bounded tasks that autonomous agents execute all day.

Built for agent busywork

Muse Ember is live today in the Muse Code editor and through the Meta Model API under the identifier muse-ember-1.0.

Meta priced it aggressively. For prompts under 200k tokens, it costs $0.10 per million input tokens and $0.50 per million output tokens. For context windows above 200k tokens, prices step up to $0.30 input and $1.50 output.

Despite its compact footprint, Ember handles text, image, audio, and video inputs natively. That means your agent can inspect bug screenshots, ingest meeting recordings, or parse invoices without needing a separate transcription or OCR model upstream.

Mark Zuckerberg noted that Ember delivers higher intelligence per dollar than others in its class. On the Artificial Analysis Intelligence Index, Ember posted a score of 36, tying Claude Haiku 5.5 and soundly beating GPT-6 Luna at 28.

The real punchline is the execution cost. Artificial Analysis found that while its token list price looks identical to Haiku 5.5, Ember solves tasks more efficiently. Running through benchmark tasks cost Ember about $0.14 per task compared to $0.41 for Haiku, effectively cutting total job costs to roughly one-third.

Ember also led its peers on AutomationBench-AA (37.8) and Terminal-Bench 4.0 (40.9%). Its main weakness is raw world knowledge: it scored just 21.1% on general question-answering, largely because it opts not to guess on 64% of questions. The upside is a hallucination rate of 17.4%, about half that of Haiku.

The architect and the workers

Meta envisions Ember acting as a fleet of cheap digital assistants orchestrated by a flagship model like Muse Wonder or Spark.

To prove the concept, Meta used Muse Code to port 16 Python lexers from Pygments into Rust. Muse Wonder acted as the supervisor: it mapped dependencies, defined task boundaries, and assigned work without ever touching the source directly. Meanwhile, 16 parallel Muse Ember workers wrote the Rust code, ran local tests, and fixed regressions.

Ember chewed through 158 million input tokens during the migration, running up a bill of just $1.84. Combined with Wonder's planning overhead, the entire multi-language migration cost under $4.

Meta ran a similar test organizing 10,000 unlabelled photos into categorized albums. Wonder wrote the sorting script and chose the taxonomy; Ember executed the bulk labeling and renaming. The entire run cost $1.33, with Ember accounting for just $0.52. Running the same volume entirely on Wonder would have easily topped $20.

Meta confirmed Ember is the first model trained using Muse Wonder Full, though it remains a closed hosted API with no weights on Hugging Face.

Why it matters

While frontier labs spend billions chasing higher reasoning ceilings, real-world agent builders keep crashing into an operational ceiling: runaway inference costs.

We think Meta's approach with Muse Ember is remarkably sharp. Real production workflows don't need a PhD-level model to run regex checks or parse receipts. By pairing an expensive frontier brain with dirt-cheap, multimodal executors, Meta provides a viable economic blueprint for running multi-agent swarms at scale.

For developers building coding harnesses or document pipelines, Ember looks like an immediate no-brainer to test. Because the API mirrors standard OpenAI endpoint formats, swapping it into an existing pipeline takes minutes.

For competitors like Anthropic and OpenAI, the pressure is on. Matching token prices on a tier list is no longer enough. If smaller models burn three times as many tokens to finish the same tool-use task, developers looking at their monthly bills will quickly migrate to where the math actually works.