Muse Ember review: ties GPT-6 Luna, but costs 6x more per task

By LLMTest Lab · Nov 10, 2026 · 4 min read musehotbenchmarksbudget-models
On this page

On this page

  1. What Meta shipped
  2. How we tested
  3. Ember vs Haiku 5.5
  4. Ember vs GPT-6 Luna
  5. The token bill
  6. Who should pick what

Muse Ember is a good cheap agent model, but in our test it was not the cheapest way to run one. It beat Claude Haiku 5.5 on 4 of 8 prompts with no losses, and it tied GPT-6 Luna on 7 of 8, losing the eighth. The catch is output volume: Ember averaged 1,141 output tokens per answer against 153 for Luna, so an average answer cost $0.00069 against $0.00011, about 6.6 times more, at the same $0.10 / $0.50 list price.

Muse Ember Claude Haiku 5.5 GPT-6 Luna
Input / output per 1M tokens $0.10 / $0.50 $0.10 / $0.50 $0.10 / $0.50
Context window 1,048,576 1,000,000 1,050,000
Pairwise wins in our test (of 16) 4 0 7
Avg cost per answer $0.00069 $0.00019 $0.00011
Avg output tokens 1,141 319 153
Avg latency 3.2 s 2.3 s 3.3 s

Prices and context sizes come from OpenRouter's model list and match Meta's launch post. Everything else is from our runs below.

What Meta shipped

Meta released Muse Ember on 2026-11-09 and calls it its smallest and most efficient model, aimed at work agents repeat all day: tool calls, coding subtasks, routing, extraction and long document parsing. It takes text, image, video, file and audio input. Meta's post says it gives more intelligence per dollar than Haiku 5.5 or Luna at any reasoning setting, citing Artificial Analysis scores of 33.9 to 35.8 for Ember against 21.4 to 27.8 for Luna. Those are vendor-reported figures, and we did not reproduce them. Our test is narrower and uses plain API calls with no reasoning setting changed.

How we tested

We wrote 8 prompts that look like jobs an agent worker gets: choosing a tool call for a refund request, extracting fields from a German invoice, routing a support ticket, counting rows in a 40-line log, normalizing messy dates and amounts, deciding whether to ask the user before a production deploy, turning a chat thread into action items, and writing a PostgreSQL query. We ran meta/muse-ember, anthropic/claude-haiku-5.5 and openai/gpt-6-luna through OpenRouter with a 1,500-token cap and default settings. The judge was anthropic/claude-sonnet-5.5, run on every pair in both answer orders. A win counts only when both orders agree; otherwise it is a tie. Total spend for the post was $0.33 of a $5 cap, on 2026-11-10. Eight prompts is a small sample, so read this as a direction. The saved runs are in the repo under server/blog/runs/muse-ember-review-2026/.

Ember vs Haiku 5.5

Ember won 4 prompts and lost none, with 4 ties. The wins were the tool-call prompt, the invoice, the log count and the deploy decision, and nearly all of them came down to format discipline. Our prompts asked for JSON only, and Haiku 5.5 wrapped its answers in markdown code fences. The judge on the tool-call prompt:

Response B wraps the array in a markdown code fence, which is extra non-JSON text and could break a strict machine parser. It also uses "tool" as the key, which is equally arbitrary.

If your pipeline strips fences, that gap disappears. If it does not, it is a real parsing failure on every call. On routing, the action-item thread and the SQL query the two were level on correctness.

Ember vs GPT-6 Luna

Luna won 1 prompt and Ember none, with 7 ties. The one loss was the SQL query. The judge's note:

Response B adds a tiebreaker (c.country) to the ORDER BY, so the result is deterministic.

On everything else the two returned the same answers, so quality did not separate them. Cost did. Luna also beat Haiku 5.5 on 6 prompts with 2 ties, so in this set it was the cleanest and cheapest of the three.

The token bill

Ember's one-line tool-call answer used 1,114 output tokens, and its hardest prompt ran to 2,985 tokens and 9.8 seconds. The visible answers were short, so the extra volume looks like reasoning that is billed as output; that is our inference from the token counts, not something Meta documents here. At $0.50 per million output tokens it still rounds to under a tenth of a cent per answer, but across a million agent steps the average gap is about $580 ($690 against $110 for Luna). If you run Ember, cap output tokens and watch the bill per task, not the price per token. The same logic drives our cost breakdown of agent loops in production.

Who should pick what

Pick GPT-6 Luna for high-volume extraction, routing and tool selection where it matches Ember's answers for a sixth of the cost. Pick Ember if your own evals show it winning on long documents, video or computer use, which our 8 prompts did not cover, and if you can bound its output. Pick Haiku 5.5 only with a prompt that forbids code fences or a parser that strips them. For a wider look at budget options see our cheap model shootout and the list of cheapest models for function calling.

Related articles

Gemini 2.5 Flash vs Haiku 4.5 vs GPT-5 mini: GPT-6 Luna beats all 3
We ran Gemini 2.5 Flash, Claude Haiku 4.5 and GPT-5 mini against two newer budget models on 8 real tasks. GPT-6 Luna won at $0.19 per 1,000 answers.
Cheapest LLMs for function calling in 2026: ranked by reliability
GPT-4o-mini, Gemini 2.5 Flash, DeepSeek V4-Flash, and Haiku 4.5 ranked on schema errors, retry costs, and multi-turn reliability for function calling in 2026.