Chat model

Qwen3.5-397B-A17B high · chat

high reasoning · 16 questions per request

Accuracy, all 2,198 questions
82.7%81.1% to 84.2%, 95% interval; 1,818 right
Answers kept at 99% confidence
78 of 2,1988 wrong; share right 81.0% to 94.7%, 95% interval
Cost, $ per 1,000 questionsper 1M questions
$2.411$2,411Basis and price under Cost
Latency
251.3 smedian request, 16 questions per request

Back to the leaderboard.

How we called it

Kind
Chat model
Access
OpenRouter chat API, DeepInfra's endpoint (fp8), fallbacks off, output cap 65,536 tokens, reasoning included
Paid as
Provider service
Model ID
qwen/qwen3.5-397b-a17b
Parameters
Mixture of experts: 397B total, about 17B active per step. Size class uses the total (397B).mixture of experts, as Qwen states it; source: huggingface.co/Qwen/Qwen3.5-397B-A17B; read 9 November 2026
Confidence from
Written estimate
Run
870, 8 November 2026

Settings

Reasoning
high
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
b0118bbae7ac
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions82.7%81.1% to 84.2%1,818 of 2,198
Reading91.0%88.6% to 93.0%597 of 656
Grammar79.2%77.1% to 81.1%1,221 of 1,542

Comparison group: mno. Compare with another setup.

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading90.8%119 of 131
Sentence meaningsReading95.5%147 of 154
How a writer explains and develops ideasReading80.0%56 of 70
Topic and main point of a paragraphReading96.9%95 of 98
How a paragraph fits togetherReading82.1%87 of 106
Details that support a paragraph’s main pointReading95.9%93 of 97
Word types, such as nouns and verbsGrammar90.0%126 of 140
Groups of words built around a nounGrammar92.3%60 of 65
Verbs and their formsGrammar86.8%79 of 91
Verb forms used like nouns or modifiersGrammar93.1%67 of 72
Word endings and what they doGrammar83.5%66 of 79
How words are builtGrammar72.5%50 of 69
The roles of words and phrases in a sentenceGrammar83.4%121 of 145
How verbs show who acts and who receives the actionGrammar91.9%68 of 74
Types of sentencesGrammar89.4%76 of 85
Questions combining several grammar topicsGrammar67.3%148 of 220
Sound patterns and changesGrammar54.5%42 of 77
SpellingGrammar70.7%94 of 133
PunctuationGrammar71.8%84 of 117
Problems in sentence wordingGrammar80.0%140 of 175

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
500 of 547, first English run (871)
Accuracy
91.4%
Non-hallucination with “I don't know”
6.2%: 3 “I don't know” of 48 not right (run 873)
Answered with “I don't know”
99.5%: 544 of 547
Error rate with “I don't know”
8.3%: 45 of 544 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
1.7421,742at list prices
Latency, median request, first English run
162.9 s; 4 requests at once
English run
871
Machine, English runs
hosted API
Date
8 November 2026
Request format
16 questions per request
Reasoning
high
Prompt version
5c3a4849752e
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

Confidence

Answers kept at 99% or more
78, of which 8 wrong (share right: 81.0% to 94.7%, 95% interval); 2,120 skipped, the model less than 99% sure
Calibration gap
6.2 points
AUROC, all questions
0.74 0.71 to 0.77
AUROC, reading
0.78 0.72 to 0.84
AUROC, grammar
0.73 0.70 to 0.76
Brier score
0.130
Accuracy of its surest answers
most confident fifth: 93.6% right; most confident half: 92.4% right; most confident four fifths: 88.8% right
Typical score of a right answer
90.0%
Typical score of a wrong answer
85.0%

Repeats

We did not repeat this setup.

Latency

Latency, median request
251.3 s (this run)
Latency, 90th percentile
386.7 s
Requests at once
4
Throughput
14.4 questions per minute of request time
Request time
152.3 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
$2.411$2,411
Whole run
$5.30
Basis
Estimated from published prices and recorded token usage
Price
$0.45 input, $0.22 cached input, $3 output per 1M tokens (source, 2026-11-08)
Input tokens per question
237
Output tokens per question
768