Chat model

Muse Ember max · chat

max reasoning · 16 questions per request

Accuracy, all 2,198 questions
85.6%84.1% to 87.0%, 95% interval; 1,882 right
Answers kept at 99% confidence
1 of 2,1980 wrong; share right 20.7% to 100.0%, 95% interval
Cost, $ per 1,000 questionsper 1M questions
$0.650$650Basis and price under Cost
Latency
30.9 smedian request, 16 questions per request

Back to the leaderboard.

How we called it

Kind
Chat model
Access
OpenRouter chat API, Meta's own endpoint, fallbacks off, output cap 65,536 tokens, reasoning max
Paid as
Provider service
Model ID
meta/muse-ember
Parameters
not disclosed by makerpages checked: research.meta.ai/blog/muse-ember, dev.meta.ai/models/muse-ember; read 10 November 2026
Note
Meta's closed chat model, released 9 November 2026, added as a chat-model comparison
Confidence from
Written estimate
Run
939, 10 November 2026

Settings

Reasoning
max
Request format
16 questions per request
Requests at once
4
Temperature
not recorded
Run record: harness seed, harness commit and prompt version
Harness seed
0
Harness commit
bb707d5f9bff
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions85.6%84.1% to 87.0%1,882 of 2,198
Reading90.2%87.7% to 92.3%592 of 656
Grammar83.7%81.7% to 85.4%1,290 of 1,542

Comparison group: m. Compare with another setup.

Invalid answers: 48; read from broken replies: 0; empty replies asked again: 0. Still invalid: 48 cut off at the output limit. An empty reply is asked once more; in a broken reply, any letter that is clearly readable is scored and the rest count as wrong. Invalid answers count as wrong (how).

With “Bilmiyorum” (run 940): Invalid answers: 48; read from broken replies: 0; empty replies asked again: 0. Still invalid: 48 cut off at the output limit. (how)

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading92.4%121 of 131
Sentence meaningsReading96.8%149 of 154
How a writer explains and develops ideasReading78.6%55 of 70
Topic and main point of a paragraphReading94.9%93 of 98
How a paragraph fits togetherReading79.2%84 of 106
Details that support a paragraph’s main pointReading92.8%90 of 97
Word types, such as nouns and verbsGrammar95.7%134 of 140
Groups of words built around a nounGrammar95.4%62 of 65
Verbs and their formsGrammar96.7%88 of 91
Verb forms used like nouns or modifiersGrammar94.4%68 of 72
Word endings and what they doGrammar93.7%74 of 79
How words are builtGrammar89.9%62 of 69
The roles of words and phrases in a sentenceGrammar91.0%132 of 145
How verbs show who acts and who receives the actionGrammar91.9%68 of 74
Types of sentencesGrammar94.1%80 of 85
Questions combining several grammar topicsGrammar80.5%177 of 220
Sound patterns and changesGrammar28.6%22 of 77
SpellingGrammar69.9%93 of 133
PunctuationGrammar76.9%90 of 117
Problems in sentence wordingGrammar80.0%140 of 175

Vision (preliminary)

Preliminary: these numbers are not final and may change. This setup is condition A of the Vision tab, which asks the 763 questions with underlined or numbered words as text (A), as a page image (B), as both (C) and through OCR models (D); the share answered right, its 95% interval in the number's tooltip.

ModelA · text onlyaccuracy, higher is betterB · image onlyaccuracy, higher is betterC · text plus imageaccuracy, higher is betterBest OCR routeaccuracy, higher is better
Muse Ember max · chat Closed · Efficient83.6%86.2%87.8%Queued3 of 4 OCR routes done; 559 of 763 asked via Baidu Unlimited-OCR-2

Calibration on these questions (ECE, Brier score, AUROC): not measured, a plain answer line gives no option probabilities, so accuracy only. Muse Ember · max · chat's condition A is its Text tab run, asked with 16 questions per request and a stated confidence; its B, C and D were asked one question per request with a plain answer line, so A differs from them in how it was asked as well as in its input.

English

The Text tab's 547 kept reading questions translated into English, asked with this setup's Text tab settings. Every setup: English reading; against Turkish: English and Turkish.

Right answers
490 of 547, first English run (942)
Accuracy
89.6%
Non-hallucination with “I don't know”
5.3%: 3 “I don't know” of 57 not right (run 943)
Answered with “I don't know”
99.5%: 544 of 547
Error rate with “I don't know”
9.9%: 54 of 544 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
0.245245at list prices
Latency, median request, first English run
11.6 s; 4 requests at once
English run
942
Machine, English runs
hosted API
Date
10 November 2026
Request format
16 questions per request
Reasoning
max
Prompt version
5c3a4849752e
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

Confidence

Answers kept at 99% or more
1, of which 0 wrong (share right: 20.7% to 100.0%, 95% interval); 2,197 skipped: 2,149 with the model less than 99% sure and 48 without a valid answer
Calibration gap
6.2 points
AUROC, all questions
0.80 0.77 to 0.83
AUROC, reading
0.86 0.81 to 0.90
AUROC, grammar
0.78 0.74 to 0.81
Brier score
0.095
Accuracy of its surest answers
most confident fifth: 97.7% right; most confident half: 96.6% right; most confident four fifths: 93.3% right
Typical score of a right answer
88.0%
Typical score of a wrong answer
65.0%

Repeats

We did not repeat this setup.

Latency

Latency, median request
30.9 s (this run)
Latency, 90th percentile
64.6 s
Requests at once
4
Throughput
102.8 questions per minute of request time
Request time
21.4 minutes, over the requests whose answers were kept

Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.

Cost

Cost per 1,000 questionsper 1M questions
$0.650$650
Whole run
$1.43
Basis
Estimated from published prices and recorded token usage
Price
$0.10 input, $0.004 cached input, $0.50 output per 1M tokens (source, 2026-11-10)
Input tokens per question
270
Output tokens per question
1,248