Decision model

SubQ Snap 1.1

1 question per request

63.9% of the 2,198 questions right (61.8% to 65.9%, 95% interval). Back to the leaderboard.

How we called it

Kind

Decision model

Access

SubQ Decisions API (public preview)

Paid as

Provider service

Model ID

subq-snap-1.1

Note

SubQ's subq-snap-1.1, in public preview, through its Decisions API, which also takes images. Its option probabilities arrive as 32-bit floats, so they sum to 1 only within 32-bit rounding. Its probability check used an experimental 32-bit rule of ours. For this model only, the check before its runs accepted a sum within one 32-bit epsilon (2^-23) per option when every value is an exact 32-bit float, instead of 7 decimal places; this is our tolerance, not an error bound SubQ states. The values are stored as SubQ returned them. It was paused twice while the approval of its probability check was confirmed, for 12 minutes and then for 3 hours 24 minutes, and resumed from its stored answers. The two requests in flight count those pauses in their times (706 and 12,236 seconds; the first after one retry, when SubQ had closed the connection), so the run's total request time and questions per minute include them. Without them, its median request time stays 0.378 s and its 90th percentile moves from 0.388 s to 0.387 s.

Confidence from

Option score

Run

346, 1 November 2026

Settings

Reasoning

off

Request format

1 question per request

Requests at once

1

Temperature

not recorded

Harness seed

0

Harness commit

0ad2797acd4d

Prompt version

254597b8be16

Results

Comparison group: stu. Compare with another setup.

Page images (preliminary)

Preliminary: these numbers are not final and may change. This setup is condition A of the Page images 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, with its 95% interval.

Confidence

Answers kept at 99% or more

299, of which 1 wrong (share right: 98.1% to 99.9%, 95% interval); 1,899 skipped, the model less than 99% sure

Calibration gap

11.2 points

AUROC, all questions

0.84 0.82 to 0.85

AUROC, reading

0.92 0.89 to 0.95

AUROC, grammar

0.77 0.75 to 0.79

Brier score

0.175

Accuracy of its surest answers

most confident fifth: 99.5% right; most confident half: 88.2% right; most confident four fifths: 72.8% right

Typical score of a right answer

93.3%

Typical score of a wrong answer

56.2%

Repeats

Run 5 times on all questions.

Answers changed in 5 runs

0 of 2,198

Same wrong answer in all 5

794 of 2,198

Wrong in at least one run

794 of 2,198

Accuracy in each run

63.9%, 63.9%, 63.9%, 63.9%, 63.9%

Latency

Latency, median request

0.378 s (this run)

Latency, 90th percentile

0.388 s

Requests at once

1

Throughput

9.6 questions per minute of request time

Request time

229.6 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

$ per 1,000 questions

$0.015

Whole run

$0.034

Basis

Estimated from published prices and recorded token usage

Price

$0.04 input, output free per 1M tokens (source, 2026-11-01)

Input tokens per question

385

Output tokens per question

0

Going first in front of an LLM

Replay of recorded answers, cutoff chosen on half the questions, averaged over the splits. Savings are estimates from API prices; the Method page explains how.