• Artificial Analysis scores Muse Ember 36 on its Intelligence Index, matching Claude Haiku 5.5 at max effort, at $0.14 per task versus $0.41.
• Ember trails Haiku 5.5 on FrontierCode 2.0 (27.2% vs 34.9%) and on AA-Omniscience accuracy (21.1% vs 34.7%).
• Terminal-Bench 4.0 numbers for Haiku 5.5 differ between Meta (39.2%) and Artificial Analysis (32.8%), so the harness matters more than the headline.
• Meta's own tables show the largest Ember leads, and Meta says each model used its own best harness; treat them as claims to reproduce.
• Route high-volume, checkable, low-stakes agent steps to Ember; keep hard coding and fact-dependent work on a stronger model.

Meta released Muse Ember on November 9, 2026, and the independent numbers back a narrower claim than the launch framing. It is a cheap, capable worker for repetitive agent steps. It is not a general replacement for Claude Haiku 5.5, and it should not be the default for hard coding or fact-dependent work. Meta says it offers more intelligence per dollar than Haiku 5.5 or GPT-6 Luna at every reasoning setting; the third-party data supports that on cost, less so on breadth.

The cost case holds up

Artificial Analysis scores Muse Ember 36 on its Intelligence Index at max effort, matching Claude Haiku 5.5, at $0.14 per task against $0.41. Per-token prices match Haiku 5.5 and GPT-6 Luna at $0.10 input and $0.50 output per million tokens, per Artificial Analysis. The saving comes from token use: Ember reportedly uses roughly a third as many output tokens, and cached input is about 60% cheaper. The larger Muse Spark 1.3 scores 37 at $2.39 per task, per The Decoder.

Coding is the clear weak spot

On FrontierCode 2.0, a test of real coding tasks, Ember scores 27.2% against 34.9% for Haiku 5.5, a result both Meta and Artificial Analysis report. Yet it leads on Terminal-Bench 4.0 and on the Coding Agent Index, 46 in Muse Code versus 37 for Haiku 5.5 in Claude Code. The split suggests Ember suits scripted, terminal-style subtasks and agent loops rather than open-ended software engineering, though the gap between benchmarks is itself a reason for testing on your own repositories.

Benchmark harnesses change the story

The same model pair produces different Terminal-Bench 4.0 results depending on who runs it. Meta lists Ember at 40.9% and Haiku 5.5 at 39.2%, while Artificial Analysis lists Haiku 5.5 at 32.8%. Meta states each model ran in its own best harness, and its page also reports internal evaluations that cannot be independently verified from the page. Treat Meta's larger margins, such as AutomationBench-AA at 37.8% versus 13.9%, as claims to reproduce rather than settled results.

It refuses often, and knows less

Artificial Analysis reports Ember's AA-Omniscience accuracy at 21.1%, the lowest among similarly capable models and well under Haiku 5.5's 34.7%. It declines 64% of questions, which holds its hallucination rate to 17.4%, about half of Haiku 5.5's. For an agent, that means fewer confident errors but more abstentions to handle. Teams should not use Ember as a knowledge source, and should build explicit fallback paths when it declines.

Who should use it, and who should not

Use Muse Ember for high-volume, checkable steps: tool-call selection, request routing, field extraction and long-document parsing, where output can be validated and a cheaper retry costs little. Skip it as the primary model for complex code generation, for answers that must be factually correct without retrieval, or where partner praise is your only evidence. Meta's cited partner quotes describe early testing, not formal evaluations. Run a small evaluation on your own traffic before shifting production routing.