---
type: "post"
id: "plotly-s-analytics-exam-puts-muse-ember-off-67701f"
slug: "plotly-s-analytics-exam-puts-muse-ember-off-67701f"
title: "Plotly's analytics exam puts Muse Ember off the cost frontier versus GPT-6 Luna"
description: "Plotly’s DataAnalyticsBench run gave Meta’s Muse Ember and GPT-6 Luna the same score, 38 of 43, but Ember’s full run cost $1.04 versus Luna’s $0.13."
retrieval_nugget: "Plotly’s DataAnalyticsBench run gave Meta’s Muse Ember and GPT-6 Luna the same score, 38 of 43, but Ember’s full run cost $1.04 versus Luna’s $0.13. The result matters to teams evaluating agent models: low token prices do not guarantee low task costs when an agent makes many more queries."
published_at: "2026-11-14"
updated_at: "2026-11-14"
record_date: "2026-11-14"
date_kind: "published_at"
topics: ["claude"]
entities: ["Plotly","Muse Ember","GPT-6 Luna","Claude Haiku 5.5","Muse Glimmer 30B","DataAnalyticsBench","Meta"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "Plotly’s benchmark found Muse Ember matched GPT-6 Luna’s score but cost $1.04 per run versus Luna’s $0.13, placing Ember off the cost frontier."
prior_refs: []
source_url: "https://plotly.com/blog/meta-muse-ember-plotly-data-analytics-bench"
source_urls: ["https://plotly.com/blog/meta-muse-ember-plotly-data-analytics-bench"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:plotly-s-analytics-exam-puts-muse-ember-off-67701f"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/plotly-s-analytics-exam-puts-muse-ember-off-67701f/","markdown":"https://newruntime.com/posts/plotly-s-analytics-exam-puts-muse-ember-off-67701f.md","json":"https://newruntime.com/posts/plotly-s-analytics-exam-puts-muse-ember-off-67701f.json"}
markdown_url: "https://newruntime.com/posts/plotly-s-analytics-exam-puts-muse-ember-off-67701f.md"
---

# Plotly's analytics exam puts Muse Ember off the cost frontier versus GPT-6 Luna

## Retrieval answer

Plotly’s DataAnalyticsBench run gave Meta’s Muse Ember and GPT-6 Luna the same score, 38 of 43, but Ember’s full run cost $1.04 versus Luna’s $0.13. The result matters to teams evaluating agent models: low token prices do not guarantee low task costs when an agent makes many more queries.

## Summary
Plotly’s DataAnalyticsBench run gave Meta’s Muse Ember and GPT-6 Luna the same score, 38 of 43, but Ember’s full run cost $1.04 versus Luna’s $0.13. The result matters to teams evaluating agent models: low token prices do not guarantee low task costs when an agent makes many more queries.

## What happened
Plotly published results from running Meta’s Muse Ember through its data analytics exam on November 10, 2026. Meta had released Ember on November 9, describing it as its smallest, most efficient Muse model for recurring agent work such as tool calls, coding subtasks, routing, extraction and long documents. Meta says the model handles text, images, audio and video and has a one-million-token context window. The weights are not public, and Meta has not disclosed its parameter count. Plotly says Ember is available through Meta’s Model API and Muse Code, as well as providers such as OpenRouter.

In Plotly’s exam, Ember scored 38 out of 43 questions, or 88%, matching GPT-6 Luna, scoring one point above Claude Haiku 5.5, and three above Meta’s open-weight Muse Glimmer 30B. Ember was perfect on 20 analytics questions, five ambiguous questions, five impossible requests and three data-quality checks. It identified four of five statistics traps, but only one of five hidden patterns. Plotly’s test harness runs models in Plotly Studio Embedded, an analytics agent for Dash apps configured for data queries, charts and multi-turn tool use.

The run cost $1.04 for Ember, compared with $0.13 for GPT-6 Luna, $0.38 for Claude Haiku 5.5 and $3.18 for Muse Glimmer 30B. Plotly reports that Ember read about 26.2 million input tokens in the detailed results section; its introduction describes the total as about 26.6 million. About 70% of input tokens came from cache, priced at $0.004 per million. Plotly lists Ember’s rates as $0.10 per million input tokens and $0.50 per million output tokens. The much larger token volume, rather than a high listed input price, drove the relative cost of this exam.

The difference came partly from how Ember handled open-ended work. On five hidden-pattern questions, it averaged 78 database queries, compared with a 25-query median across the 45 models on Plotly’s board; two questions required more than 100 queries. On questions with a clear answer, its median was nine queries. Ember’s median time per question was 28 seconds, near the board’s 27-second median. Plotly says the model spent about 24 minutes generating across the exam; one question took 75 minutes end to end, mostly because the database ran a heavy query.

Plotly describes the exam as a private test using a synthetic wind-energy fleet warehouse, to keep the answer key out of training data and avoid answers existing elsewhere. It says the same harness and grading were used for each model and that the exam is shared with researchers. Plotly cautions that its reported Ember result is one run; repeat runs of other models have varied by as much as three points.

## What the archive adds
New Runtime covered Muse Ember’s launch on [November 9](https://newruntime.com/posts/meta-launched-muse-ember-its-smallest-model-built-4baf24/) and its pricing comparison on [November 10](https://newruntime.com/posts/meta-introduced-muse-ember-1-0-a-reasoning-1cdb9e/). Plotly’s result is a new fact on that story: a task-level benchmark where Ember tied GPT-6 Luna’s score but cost substantially more in a data-analysis workflow.

## Trend signal
This result supports the shift toward evaluating agent economics by completed-task cost rather than token price. Ember’s nominal rates did not predict its bill in Plotly’s test because its query behavior drove much higher token use. The evidence is specific to one benchmark, but it makes the distinction concrete: [Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/).

## Why now
Plotly says it runs newly released models through the same exam and shares the results as they appear. This run follows Meta’s November 9 release and tests the company’s cost-performance positioning against the actual token and query volume of an analytics task.

## Why it matters
For teams building agent workflows, the useful budget unit is often a completed task, not a million tokens. A model can have attractive per-token pricing and still spend more overall if the harness lets it explore broadly, repeat queries or struggle to identify which dimension explains a result. Plotly’s test also shows why score alone is insufficient: Ember tied Luna overall, but missed four of five hidden patterns, a category that can matter when analysts rely on agents to surface less obvious anomalies.

The findings are not a universal ranking. They are a reason to test candidate models with representative data, task definitions and tool constraints, while logging total input and output tokens, tool calls, elapsed time and correctness. Teams should compare the cost of a verified outcome and inspect failure categories before making a model choice based on headline rates or aggregate accuracy.

## Limits and uncertainty
Plotly reports one Ember run, so the score and cost are not a statistical estimate of repeated performance. It says other models’ scores have moved by up to three points across repeat runs; no repeat-run range for Ember is reported here. The exam uses a private synthetic wind-energy dataset and a Plotly analytics harness, so results may not transfer to real company data, other tools or different instructions. The source also gives two slightly different totals for Ember’s tokens read: about 26.6 million in the introduction and about 26.2 million in the detailed section. The test does not establish how Ember compares across other workloads.

## Reader takeaway
Muse Ember matched GPT-6 Luna’s score on Plotly’s analytics exam, but its much heavier query and token use made that run eight times more expensive. If you are choosing an agent model, benchmark it on your own tasks and compare cost per correct, verified result—not just the posted price per token.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)