Chat model

Gemini 4 Carbon high · chat

high reasoning · 16 questions per request

Accuracy, all 2,198 questions
98.5%97.8% to 98.9%, 95% interval; 2,164 right
Answers kept at 99% confidence
1,797 of 2,1985 wrong; share right 99.4% to 99.9%, 95% interval
Cost, $ per 1,000 questionsper 1M questions
$31.309$31,309Basis and price under Cost
Latency
217.7 smedian request, 16 questions per request

Back to the leaderboard.

How we called it

Kind
Chat model
Access
Google Antigravity through Can's agbench bridge (Google AI Ultra subscription), one request at a time, thinking high; each request a new Antigravity agent conversation, so Antigravity's own agent instructions and tool declarations (about 13,000 tokens) come before our prompt, with every tool call denied and, as the bridge's owner reports, no workspace files and no earlier conversation; the bridge applies no response format, so the prompt adds a paragraph stating the answer layout
Paid as
Subscription
Model ID
gemini-4-carbon
Parameters
not disclosed by makerpages checked: deepmind.google/models/model-cards/gemini-4-carbon/; read 11 November 2026
Note
One request of 16 questions came back empty twice during the subscription bridge's outage on 14 November 2026; it was asked a third time on 15 November 2026 at the owner's decision, with the same settings, and answered (15 of the 16 right). Both empty replies are kept in the run's record.
Confidence from
Written estimate
Run
1108, 14 November 2026

Settings

Reasoning
high
Request format
16 questions per request
Requests at once
1
Temperature
not recorded
Run record: harness seed, harness commit and prompt version
Harness seed
0
Harness commit
8eedc6a1394c
Prompt version
254597b8be16

Results

QuestionsAccuracy95% intervalRight
All questions98.5%97.8% to 98.9%2,164 of 2,198
Reading97.6%96.1% to 98.5%640 of 656
Grammar98.8%98.2% to 99.3%1,524 of 1,542

Comparison group: a. Compare with another setup.

Invalid answers: 0; read from broken replies: 0; empty replies asked again: 4 (4 came back with answers, 64 answers). Without the rule: 64 invalid. 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 1121): Invalid answers: 0; read from broken replies: 0; empty replies asked again: 4 (4 came back with answers, 64 answers). Without the rule: 64 invalid. (how)

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading96.2%126 of 131
Sentence meaningsReading99.4%153 of 154
How a writer explains and develops ideasReading97.1%68 of 70
Topic and main point of a paragraphReading95.9%94 of 98
How a paragraph fits togetherReading97.2%103 of 106
Details that support a paragraph’s main pointReading99.0%96 of 97
Word types, such as nouns and verbsGrammar99.3%139 of 140
Groups of words built around a nounGrammar100.0%65 of 65
Verbs and their formsGrammar100.0%91 of 91
Verb forms used like nouns or modifiersGrammar98.6%71 of 72
Word endings and what they doGrammar98.7%78 of 79
How words are builtGrammar97.1%67 of 69
The roles of words and phrases in a sentenceGrammar98.6%143 of 145
How verbs show who acts and who receives the actionGrammar98.6%73 of 74
Types of sentencesGrammar98.8%84 of 85
Questions combining several grammar topicsGrammar99.1%218 of 220
Sound patterns and changesGrammar98.7%76 of 77
SpellingGrammar99.2%132 of 133
PunctuationGrammar98.3%115 of 117
Problems in sentence wordingGrammar98.3%172 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
516 of 547, first English run (1055)
Accuracy
94.3%
Non-hallucination with “I don't know”
25.0%: 4 “I don't know” of 16 not right (run 1067)
Answered with “I don't know”
99.3%: 543 of 547
Error rate with “I don't know”
2.2%: 12 of 543 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
38.05338,053subscription, priced as the API
Latency, median request, first English run
287.8 s; 1 request at once
English run
1055
Machine, English runs
hosted API
Date
12 November 2026
Request format
16 questions per request
Reasoning
high
Prompt version
5c3a4849752e
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

Confidence

Answers kept at 99% or more
1,797, of which 5 wrong (share right: 99.4% to 99.9%, 95% interval); 401 skipped, the model less than 99% sure
Calibration gap
0.4 points
AUROC, all questions
0.89 0.82 to 0.95
AUROC, reading
0.93 0.84 to 0.98
AUROC, grammar
0.87 0.76 to 0.95
Brier score
0.013
Accuracy of its surest answers
most confident fifth: 99.8% right; most confident half: 99.7% right; most confident four fifths: 99.7% right
Typical score of a right answer
99.0%
Typical score of a wrong answer
93.0%

Repeats

We did not repeat this setup.

Latency

Latency, median request
217.7 s (this run)
Latency, 90th percentile
392.1 s
Requests at once
1
Throughput
3.6 questions per minute of request time
Request time
608.0 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
$31.309$31,309
Whole run
$68.82
Basis
Subscription run priced as if we had used the API
Price
$4 input, $0.20 cached input, $20 output per 1M tokens (source, 2026-11-12)
Input tokens per question
1,063
Output tokens per question
1,353