We send exactly the same task prompt and the same materials to the 4 AIs on the same day. We open a new chat every time so earlier answers don't carry over. The sending order is rotated in each run. We also record the seconds from sending until the answer stops.

Why do you send the same task 3 times?

Because AI answers vary even for the same question. From a single try, you can't tell whether a result reflects real ability or chance. We send each task 3 times, and the share of tasks whose conclusion matched is reported as the consistency rate.

Who does the scoring?

Scoring is done only by a program. We do not let AI score the answers. The correct answers are set when the tasks are created, and whether an answer is right is decided by whether numbers match, whether required words are present, whether the format fits, whether URLs actually open and the result of running formulas. Matters of taste are not scored.

Is the data in the tasks from real companies?

No. We create new data for a fictional company every time. We do not use real companies, real people or text from other companies. We share the materials we used as they are, so you can repeat the same test.

How do you count cases where the AI couldn't answer?

Refusals and timeouts stay in the base and are counted as wrong. If we hit the plan's usage limit or a human verification screen appears, we remove that trial from the base and treat it as not measured. Operating errors on AInformation's side are also removed from the base, and we do not report them as the AI's fault.

Do you use the free or paid versions?

We use each AI's paid plan through the browser interface. We don't use the API. For each run, we record the plan name, the model name shown on screen and the time we checked it.

Does this ranking show which AI is the smartest?

No. We measure only 5 things on a fixed set of work tasks: accuracy, time to answer, consistency rate, real-source rate and instruction compliance. We do not measure writing style, readability or behavior in long conversations. If the tasks change, the ranking changes too.

Reading numbers from images of receipts, charts, and tables (4 AIs, 6 tasks, 3 runs each)

Name: Reading numbers from images of receipts, charts, and tables (4 AIs, 6 tasks, 3 runs each)
Creator: AInformation
Published: 2026-11-03
License: https://creativecommons.org/licenses/by/4.0/

RESULT

What this run found

The line below is generated directly from the scoring results. The page adds no words of its own.

Reading images and photos: 6 tasks, 3 trials each We gave 6 tasks to AIs, 3 times each; Claude had the highest accuracy at 100.0% (18/18 trials), with ChatGPT second at 72.2% 72 valid trials. Results from a single round of testing. Rankings may change on a different day or with different tasks. See scores by AI

SCORES

Scores by AI

Sorted by accuracy, highest first. The base is the number of trials in which that metric could be measured. “—” means not measured; it does not mean 0%.

Reading numbers from images of receipts, charts, and tables (4 AIs, 6 tasks, 3 runs each): scores by AI
AI	AccuracyCorrect ÷ valid trials	ConsistencySame conclusion in all 3 trials	Real-source rateShare of URLs that opened	Instruction complianceMet length and format rules	Time to answerMedian / range
1Claude	100% 18 of 18 trials	100% out of 6 tasks	—	16.7%	13.9 s 13.8–17.1 seconds
2ChatGPT	72.2% 13 of 18 trials	83.3% out of 6 tasks	—	100%	25.4 s 20.4–48.3 seconds
3Gemini	38.9% 7 of 18 trials	100% out of 6 tasks	—	62.5%	69.2 s 19–204.9 seconds
4Copilot	0% 0 of 18 trials	100% out of 6 tasks	—	50%	14 s 13.8–17.1 seconds

The real-source rate is measured only in runs that included a task asking for URLs along with the answer. For Claude, ChatGPT, Gemini and Copilot, “—” means this run did not include that kind of task. It does not mean 0%.

Trials with no answer
ChatGPT … Timed out: 4 times · Gemini … Timed out: 10 times
Refusals and timeouts stay in the base and count as incorrect. Plan limits, human verification screens and operating errors on AInformation's side are excluded from the base.

TASKS

Results by task

We sent each task to each of the 4 AIs 3 times. ○ means correct and × means incorrect. The numbers below are the seconds each trial took. Open each section to read the full task prompt, the answer key set in advance and why each verdict was given.

Easy

Read 3 numbers from the attached receipt image. Rules - The answer must be exactly 3 lines, in this order: Subtotal: N /… Scoring type: exact_number — Whether the number matches the correct answer. The rounding tolerance is set in advance

Claude
- Claude, trial 1: Correct 13.9 s
- Claude, trial 2: Correct 13.9 s
- Claude, trial 3: Correct 13.9 s
3 of 3 trials correct
ChatGPT
- ChatGPT, trial 1: Correct 35.8 s
- ChatGPT, trial 2: Correct 29.7 s
- ChatGPT, trial 3: Correct 20.5 s
3 of 3 trials correct
Gemini
- Gemini, trial 1: Incorrect 19.2 s
- Gemini, trial 2: Incorrect Timed out
- Gemini, trial 3: Incorrect Timed out
0 of 3 trials correct
Copilot
- Copilot, trial 1: Incorrect 14 s
- Copilot, trial 2: Incorrect 13.9 s
- Copilot, trial 3: Incorrect 13.9 s
0 of 3 trials correct

Full task promptWe sent the original Japanese version of this prompt (103 Japanese characters) unchanged to all 4 AIs. The text shown here is our translation.

Read 3 numbers from the attached receipt image. Rules - The answer must be exactly 3 lines, in this order: Subtotal: N / Consumption tax: N / Total: N - Write numbers as plain digits. No yen sign or commas needed. Do not add any explanation

Answer key set in advanceWe set it when we created the task. 2 of the 4 AIs got it right at least once

(Machine-graded exact_number) Subtotal = 2,316 (margin ±0) / Consumption tax = 185 (margin ±0) / Total = 2,501 (margin ±0)

Scoring detailsWe show, trial by trial, why each verdict was given (6 of 12 trials correct)

Claude, trial 1: Correct: All figures match: Subtotal ○ (correct answer 2,316) / Consumption tax ○ (correct answer 185) / Total ○ (correct answer 2,501) (13.9 seconds) · Full text of this answer (Japanese)
Claude, trial 2: Correct: All figures match: Subtotal ○ (correct answer 2,316) / Consumption tax ○ (correct answer 185) / Total ○ (correct answer 2,501) (13.9 seconds) · Full text of this answer (Japanese)
Claude, trial 3: Correct: All figures match: Subtotal ○ (correct answer 2,316) / Consumption tax ○ (correct answer 185) / Total ○ (correct answer 2,501) (13.9 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 1: Correct: All figures match: Subtotal ○ (correct answer 2,316) / Consumption tax ○ (correct answer 185) / Total ○ (correct answer 2,501) (35.8 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 2: Correct: All figures match: Subtotal ○ (correct answer 2,316) / Consumption tax ○ (correct answer 185) / Total ○ (correct answer 2,501) (29.7 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 3: Correct: All figures match: Subtotal ○ (correct answer 2,316) / Consumption tax ○ (correct answer 185) / Total ○ (correct answer 2,501) (20.5 seconds) · Full text of this answer (Japanese)
Gemini, trial 1: Incorrect: Figures are wrong: Subtotal × (correct answer 2,316; no “Subtotal” row) / Consumption tax × (correct answer 185; no “Tax” row) / Total × (correct answer 2,501; no “Total” row) (19.2 seconds) · Full text of this answer (Japanese)
Gemini, trial 2: Incorrect: Timed out (the answer did not finish before the time limit) (302.6 seconds)
Gemini, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (302.2 seconds)
Copilot, trial 1: Incorrect: Figures are wrong: Subtotal × (correct answer 2,316; no “Subtotal” row) / Consumption tax × (correct answer 185; no “Tax” row) / Total × (correct answer 2,501; no “Total” row) (14 seconds) · Full text of this answer (Japanese)
Copilot, trial 2: Incorrect: Figures are wrong: Subtotal × (correct answer 2,316; no “Subtotal” row) / Consumption tax × (correct answer 185; no “Tax” row) / Total × (correct answer 2,501; no “Total” row) (13.9 seconds) · Full text of this answer (Japanese)
Copilot, trial 3: Incorrect: Figures are wrong: Subtotal × (correct answer 2,316; no “Subtotal” row) / Consumption tax × (correct answer 185; no “Tax” row) / Total × (correct answer 2,501; no “Total” row) (13.9 seconds) · Full text of this answer (Japanese)

Easy

Read the change amount from the attached receipt image. Rules - The answer must be exactly 1 line, in the form Change: N… Scoring type: exact_number — Whether the number matches the correct answer. The rounding tolerance is set in advance

Claude
- Claude, trial 1: Correct 14.1 s
- Claude, trial 2: Correct 13.9 s
- Claude, trial 3: Correct 13.9 s
3 of 3 trials correct
ChatGPT
- ChatGPT, trial 1: Correct 24.1 s
- ChatGPT, trial 2: Correct 20.4 s
- ChatGPT, trial 3: Correct 20.4 s
3 of 3 trials correct
Gemini
- Gemini, trial 1: Correct 59.7 s
- Gemini, trial 2: Incorrect Timed out
- Gemini, trial 3: Incorrect Timed out
1 of 3 trials correct
Copilot
- Copilot, trial 1: Incorrect 14.1 s
- Copilot, trial 2: Incorrect 14.1 s
- Copilot, trial 3: Incorrect 13.8 s
0 of 3 trials correct

Full task promptWe sent the original Japanese version of this prompt (77 Japanese characters) unchanged to all 4 AIs. The text shown here is our translation.

Read the change amount from the attached receipt image. Rules - The answer must be exactly 1 line, in the form Change: N - Write numbers as plain digits. Do not add any explanation

Answer key set in advanceWe set it when we created the task. 3 of the 4 AIs got it right at least once

(Machine-graded exact_number) Change = 499 (margin ±0)

Scoring detailsWe show, trial by trial, why each verdict was given (7 of 12 trials correct)

Claude, trial 1: Correct: All figures match: Change ○ (correct answer 499) (14.1 seconds) · Full text of this answer (Japanese)
Claude, trial 2: Correct: All figures match: Change ○ (correct answer 499) (13.9 seconds) · Full text of this answer (Japanese)
Claude, trial 3: Correct: All figures match: Change ○ (correct answer 499) (13.9 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 1: Correct: All figures match: Change ○ (correct answer 499) (24.1 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 2: Correct: All figures match: Change ○ (correct answer 499) (20.4 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 3: Correct: All figures match: Change ○ (correct answer 499) (20.4 seconds) · Full text of this answer (Japanese)
Gemini, trial 1: Correct: All figures match: Change ○ (correct answer 499) (59.7 seconds) · Full text of this answer (Japanese)
Gemini, trial 2: Incorrect: Timed out (the answer did not finish before the time limit) (302.3 seconds)
Gemini, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (302.1 seconds)
Copilot, trial 1: Incorrect: Figures are wrong: Change × (correct answer 499; no “Change” row) (14.1 seconds) · Full text of this answer (Japanese)
Copilot, trial 2: Incorrect: Figures are wrong: Change × (correct answer 499; no “Change” row) (14.1 seconds) · Full text of this answer (Japanese)
Copilot, trial 3: Incorrect: Figures are wrong: Change × (correct answer 499; no “Change” row) (13.8 seconds) · Full text of this answer (Japanese)

Medium

List every item purchased, based on the attached receipt image. Rules - One item per line. Start each line with a bullet… Scoring type: set_match — How many of the required items were missed. Counts both omissions and made-up additions

Claude
- Claude, trial 1: Correct 17 s
- Claude, trial 2: Correct 13.9 s
- Claude, trial 3: Correct 13.8 s
3 of 3 trials correct
ChatGPT
- ChatGPT, trial 1: Correct 27.2 s
- ChatGPT, trial 2: Correct 32.8 s
- ChatGPT, trial 3: Incorrect Timed out
2 of 3 trials correct
Gemini
- Gemini, trial 1: Correct 189.8 s
- Gemini, trial 2: Correct 19 s
- Gemini, trial 3: Incorrect Timed out
2 of 3 trials correct
Copilot
- Copilot, trial 1: Incorrect 14.1 s
- Copilot, trial 2: Incorrect 13.9 s
- Copilot, trial 3: Incorrect 13.9 s
0 of 3 trials correct

Full task promptWe sent the original Japanese version of this prompt (93 Japanese characters) unchanged to all 4 AIs. The text shown here is our translation.

List every item purchased, based on the attached receipt image. Rules - One item per line. Start each line with a bullet ( · ). Write the item name and the price - Do not include the subtotal, sales tax, total, amount received, or change lines

Answer key set in advanceWe set it when we created the task. 3 of the 4 AIs got it right at least once

(Machine-graded set_match) Milk 1000ml [Milk +314] / Yogurt 400g [Yogurt +184] / Udon 3 servings [Udon +184] / Chicken breast 2 pieces [Chicken breast +454] / Detergent refill pack [Detergent +496] / Coffee 200g [Coffee +684] — up to 0 extra items allowed

Scoring detailsWe show, trial by trial, why each verdict was given (7 of 12 trials correct)

Claude, trial 1: Correct: All 6 items match, 0 extra items added (17 seconds) · Full text of this answer (Japanese)
Claude, trial 2: Correct: All 6 items match, 0 extra items added (13.9 seconds) · Full text of this answer (Japanese)
Claude, trial 3: Correct: All 6 items match, 0 extra items added (13.8 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 1: Correct: All 6 items match, 0 extra items added (27.2 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 2: Correct: All 6 items match, 0 extra items added (32.8 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (300.6 seconds)
Gemini, trial 1: Correct: All 6 items match, 0 extra items added (189.8 seconds) · Full text of this answer (Japanese)
Gemini, trial 2: Correct: All 6 items match, 0 extra items added (19 seconds) · Full text of this answer (Japanese)
Gemini, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (302.1 seconds)
Copilot, trial 1: Incorrect: 0 of 6 items match; missing Milk 1000ml / Yogurt 400g / Udon 3 servings / Chicken breast 2 pieces / Detergent refill pack / Coffee 200g; 1 extra item added (14.1 seconds) · Full text of this answer (Japanese)
Copilot, trial 2: Incorrect: 0 of 6 items match; missing Milk 1000ml / Yogurt 400g / Udon 3 servings / Chicken breast 2 pieces / Detergent refill pack / Coffee 200g; 1 extra item added (13.9 seconds) · Full text of this answer (Japanese)
Copilot, trial 3: Incorrect: 0 of 6 items match; missing Milk 1000ml / Yogurt 400g / Udon 3 servings / Chicken breast 2 pieces / Detergent refill pack / Coffee 200g; 1 extra item added (13.9 seconds) · Full text of this answer (Japanese)

Medium

Looking at the attached bar chart image, which month has the tallest bar? Rules - The answer must be exactly 1 line: jus… Scoring type: contains_all — Whether all the required words and phrases are included

Claude
- Claude, trial 1: Correct 13.9 s
- Claude, trial 2: Correct 13.8 s
- Claude, trial 3: Correct 13.9 s
3 of 3 trials correct
ChatGPT
- ChatGPT, trial 1: Correct 26.7 s
- ChatGPT, trial 2: Correct 48.3 s
- ChatGPT, trial 3: Incorrect Timed out
2 of 3 trials correct
Gemini
- Gemini, trial 1: Correct 61.7 s
- Gemini, trial 2: Incorrect Timed out
- Gemini, trial 3: Correct 76.7 s
2 of 3 trials correct
Copilot
- Copilot, trial 1: Incorrect 13.9 s
- Copilot, trial 2: Incorrect 14 s
- Copilot, trial 3: Incorrect 14 s
0 of 3 trials correct

Full task promptWe sent the original Japanese version of this prompt (64 Japanese characters) unchanged to all 4 AIs. The text shown here is our translation.

Looking at the attached bar chart image, which month has the tallest bar? Rules - The answer must be exactly 1 line: just the month name (for example, January) - Do not include numeric values or any explanation

Answer key set in advanceWe set it when we created the task. 3 of the 4 AIs got it right at least once

(Machine-graded contains_all) Highest month = June

Scoring detailsWe show, trial by trial, why each verdict was given (7 of 12 trials correct)

Claude, trial 1: Correct: The one item was present (13.9 seconds) · Full text of this answer (Japanese)
Claude, trial 2: Correct: The one item was present (13.8 seconds) · Full text of this answer (Japanese)
Claude, trial 3: Correct: The one item was present (13.9 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 1: Correct: The one item was present (26.7 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 2: Correct: The one item was present (48.3 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (300.8 seconds)
Gemini, trial 1: Correct: The one item was present (61.7 seconds) · Full text of this answer (Japanese)
Gemini, trial 2: Incorrect: Timed out (the answer did not finish before the time limit) (302.2 seconds)
Gemini, trial 3: Correct: The one item was present (76.7 seconds) · Full text of this answer (Japanese)
Copilot, trial 1: Incorrect: 0 of 1 item present; missing “highest month” (13.9 seconds) · Full text of this answer (Japanese)
Copilot, trial 2: Incorrect: 0 of 1 item present; missing “highest month” (14 seconds) · Full text of this answer (Japanese)
Copilot, trial 3: Incorrect: 0 of 1 item present; missing “highest month” (14 seconds) · Full text of this answer (Japanese)

Hard

From the attached bar chart image, read off every monthly value. Rules - One line per month, in the format Month: Value,… Scoring type: exact_number — Whether the number matches the correct answer. The rounding tolerance is set in advance

Claude
- Claude, trial 1: Correct 13.9 s
- Claude, trial 2: Correct 13.9 s
- Claude, trial 3: Correct 13.9 s
3 of 3 trials correct
ChatGPT
- ChatGPT, trial 1: Correct 20.5 s
- ChatGPT, trial 2: Correct 26.6 s
- ChatGPT, trial 3: Incorrect Timed out
2 of 3 trials correct
Gemini
- Gemini, trial 1: Correct 204.9 s
- Gemini, trial 2: Incorrect Timed out
- Gemini, trial 3: Incorrect Timed out
1 of 3 trials correct
Copilot
- Copilot, trial 1: Incorrect 14 s
- Copilot, trial 2: Incorrect 13.9 s
- Copilot, trial 3: Incorrect 13.9 s
0 of 3 trials correct

Full task promptWe sent the original Japanese version of this prompt (85 Japanese characters) unchanged to all 4 AIs. The text shown here is our translation.

From the attached bar chart image, read off every monthly value. Rules - One line per month, in the format Month: Value, listed from left to right - Write numbers using standard digits. Do not add any explanation

Answer key set in advanceWe set it when we created the task. 3 of the 4 AIs got it right at least once

(Machine-graded exact_number) January = 5,500 (margin ±0) / February = 5,500 (margin ±0) / March = 8,200 (margin ±0) / April = 12,300 (margin ±0) / May = 17,800 (margin ±0) / June = 24,700 (margin ±0)

Scoring detailsWe show, trial by trial, why each verdict was given (6 of 12 trials correct)

Claude, trial 1: Correct: All figures match: January ○ (correct answer 5,500) / February ○ (correct answer 5,500) / March ○ (correct answer 8,200) / April ○ (correct answer 12,300) / May ○ (correct answer 17,800) / June ○ (correct answer 24,700) (13.9 seconds) · Full text of this answer (Japanese)
Claude, trial 2: Correct: All figures match: January ○ (correct answer 5,500) / February ○ (correct answer 5,500) / March ○ (correct answer 8,200) / April ○ (correct answer 12,300) / May ○ (correct answer 17,800) / June ○ (correct answer 24,700) (13.9 seconds) · Full text of this answer (Japanese)
Claude, trial 3: Correct: All figures match: January ○ (correct answer 5,500) / February ○ (correct answer 5,500) / March ○ (correct answer 8,200) / April ○ (correct answer 12,300) / May ○ (correct answer 17,800) / June ○ (correct answer 24,700) (13.9 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 1: Correct: All figures match: January ○ (correct answer 5,500) / February ○ (correct answer 5,500) / March ○ (correct answer 8,200) / April ○ (correct answer 12,300) / May ○ (correct answer 17,800) / June ○ (correct answer 24,700) (20.5 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 2: Correct: All figures match: January ○ (correct answer 5,500) / February ○ (correct answer 5,500) / March ○ (correct answer 8,200) / April ○ (correct answer 12,300) / May ○ (correct answer 17,800) / June ○ (correct answer 24,700) (26.6 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (300.8 seconds)
Gemini, trial 1: Correct: All figures match: January ○ (correct answer 5,500) / February ○ (correct answer 5,500) / March ○ (correct answer 8,200) / April ○ (correct answer 12,300) / May ○ (correct answer 17,800) / June ○ (correct answer 24,700) (204.9 seconds) · Full text of this answer (Japanese)
Gemini, trial 2: Incorrect: Timed out (the answer did not finish before the time limit) (302.1 seconds)
Gemini, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (302.3 seconds)
Copilot, trial 1: Incorrect: Figures are wrong: January × (correct answer 5,500; no “January” row) / February × (correct answer 5,500; no “February” row) / March × (correct answer 8,200; no “March” row) / April × (correct answer 12,300; no “April” row) / May × (correct answer 17,800; no “May” row) / June × (correct answer 24,700; no “June” row) (14 seconds) · Full text of this answer (Japanese)
Copilot, trial 2: Incorrect: Figures are wrong: January × (correct answer 5,500; no “January” row) / February × (correct answer 5,500; no “February” row) / March × (correct answer 8,200; no “March” row) / April × (correct answer 12,300; no “April” row) / May × (correct answer 17,800; no “May” row) / June × (correct answer 24,700; no “June” row) (13.9 seconds) · Full text of this answer (Japanese)
Copilot, trial 3: Incorrect: Figures are wrong: January × (correct answer 5,500; no “January” row) / February × (correct answer 5,500; no “February” row) / March × (correct answer 8,200; no “March” row) / April × (correct answer 12,300; no “April” row) / May × (correct answer 17,800; no “May” row) / June × (correct answer 24,700; no “June” row) (13.9 seconds) · Full text of this answer (Japanese)

Hard

Read all the values in the Product A column from the attached table image (which is slightly tilted). Rules - One value … Scoring type: exact_number — Whether the number matches the correct answer. The rounding tolerance is set in advance

Claude
- Claude, trial 1: Correct 17.1 s
- Claude, trial 2: Correct 17 s
- Claude, trial 3: Correct 16.9 s
3 of 3 trials correct
ChatGPT
- ChatGPT, trial 1: Incorrect 20.6 s
- ChatGPT, trial 2: Correct 20.5 s
- ChatGPT, trial 3: Incorrect Timed out
1 of 3 trials correct
Gemini
- Gemini, trial 1: Correct 134.9 s
- Gemini, trial 2: Incorrect Timed out
- Gemini, trial 3: Incorrect Timed out
1 of 3 trials correct
Copilot
- Copilot, trial 1: Incorrect 14 s
- Copilot, trial 2: Incorrect 17.1 s
- Copilot, trial 3: Incorrect 14 s
0 of 3 trials correct

Full task promptWe sent the original Japanese version of this prompt (92 Japanese characters) unchanged to all 4 AIs. The text shown here is our translation.

Read all the values in the Product A column from the attached table image (which is slightly tilted). Rules - One value per line, in the form Month: N, from top to bottom - Write numbers as plain digits. Do not add any explanation

Answer key set in advanceWe set it when we created the task. 3 of the 4 AIs got it right at least once

(Machine-graded exact_number) 2025-07 = 32,900 (margin ±0) / 2025-08 = 35,600 (margin ±0) / 2025-09 = 23,300 (margin ±0) / 2025-10 = 11,000 (margin ±0)

Scoring detailsWe show, trial by trial, why each verdict was given (5 of 12 trials correct)

Claude, trial 1: Correct: All figures match: 2025-07 ○ (correct answer 32,900) / 2025-08 ○ (correct answer 35,600) / 2025-09 ○ (correct answer 23,300) / 2025-10 ○ (correct answer 11,000) (17.1 seconds) · Full text of this answer (Japanese)
Claude, trial 2: Correct: All figures match: 2025-07 ○ (correct answer 32,900) / 2025-08 ○ (correct answer 35,600) / 2025-09 ○ (correct answer 23,300) / 2025-10 ○ (correct answer 11,000) (17 seconds) · Full text of this answer (Japanese)
Claude, trial 3: Correct: All figures match: 2025-07 ○ (correct answer 32,900) / 2025-08 ○ (correct answer 35,600) / 2025-09 ○ (correct answer 23,300) / 2025-10 ○ (correct answer 11,000) (16.9 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 1: Incorrect: Figures are wrong: 2025-07 × (correct answer 32,900; no “2025-07” row) / 2025-08 × (correct answer 35,600; no “2025-08” row) / 2025-09 × (correct answer 23,300; no “2025-09” row) / 2025-10 × (correct answer 11,000; no “2025-10” row) (20.6 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 2: Correct: All figures match: 2025-07 ○ (correct answer 32,900) / 2025-08 ○ (correct answer 35,600) / 2025-09 ○ (correct answer 23,300) / 2025-10 ○ (correct answer 11,000) (20.5 seconds) · Full text of this answer (Japanese)
ChatGPT, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (301.2 seconds)
Gemini, trial 1: Correct: All figures match: 2025-07 ○ (correct answer 32,900) / 2025-08 ○ (correct answer 35,600) / 2025-09 ○ (correct answer 23,300) / 2025-10 ○ (correct answer 11,000) (134.9 seconds) · Full text of this answer (Japanese)
Gemini, trial 2: Incorrect: Timed out (the answer did not finish before the time limit) (302.1 seconds)
Gemini, trial 3: Incorrect: Timed out (the answer did not finish before the time limit) (302.5 seconds)
Copilot, trial 1: Incorrect: Figures are wrong: 2025-07 × (correct answer 32,900; no “2025-07” row) / 2025-08 × (correct answer 35,600; no “2025-08” row) / 2025-09 × (correct answer 23,300; no “2025-09” row) / 2025-10 × (correct answer 11,000; no “2025-10” row) (14 seconds) · Full text of this answer (Japanese)
Copilot, trial 2: Incorrect: Figures are wrong: 2025-07 × (correct answer 32,900; no “2025-07” row) / 2025-08 × (correct answer 35,600; no “2025-08” row) / 2025-09 × (correct answer 23,300; no “2025-09” row) / 2025-10 × (correct answer 11,000; no “2025-10” row) (17.1 seconds) · Full text of this answer (Japanese)
Copilot, trial 3: Incorrect: Figures are wrong: 2025-07 × (correct answer 32,900; no “2025-07” row) / 2025-08 × (correct answer 35,600; no “2025-08” row) / 2025-09 × (correct answer 23,300; no “2025-09” row) / 2025-10 × (correct answer 11,000; no “2025-10” row) (14 seconds) · Full text of this answer (Japanese)

ANSWERS

Full answers from the 4 AIs

These are the answers we scored, posted as is without edits. That way, anyone can recount the results the same way we did.

Claude

18 answers18 of them correct

t1, trial 1

Full text file (Japanese) · Correct

t1, trial 2

Full text file (Japanese) · Correct

t1, trial 3

Full text file (Japanese) · Correct

t2, trial 1

Full text file (Japanese) · Correct

t2, trial 2

Full text file (Japanese) · Correct

t2, trial 3

Full text file (Japanese) · Correct

t3, trial 1

Full text file (Japanese) · Correct

t3, trial 2

Full text file (Japanese) · Correct

t3, trial 3

Full text file (Japanese) · Correct

t4, trial 1

Full text file (Japanese) · Correct

t4, trial 2

Full text file (Japanese) · Correct

t4, trial 3

Full text file (Japanese) · Correct

t5, trial 1

Full text file (Japanese) · Correct

t5, trial 2

Full text file (Japanese) · Correct

t5, trial 3

Full text file (Japanese) · Correct

t6, trial 1

Full text file (Japanese) · Correct

t6, trial 2

Full text file (Japanese) · Correct

t6, trial 3

Full text file (Japanese) · Correct
ChatGPT

14 answers13 of them correct

t1, trial 1

Full text file (Japanese) · Correct

t1, trial 2

Full text file (Japanese) · Correct

t1, trial 3

Full text file (Japanese) · Correct

t2, trial 1

Full text file (Japanese) · Correct

t2, trial 2

Full text file (Japanese) · Correct

t2, trial 3

Full text file (Japanese) · Correct

t3, trial 1

Full text file (Japanese) · Correct

t3, trial 2

Full text file (Japanese) · Correct

t4, trial 1

Full text file (Japanese) · Correct

t4, trial 2

Full text file (Japanese) · Correct

t5, trial 1

Full text file (Japanese) · Correct

t5, trial 2

Full text file (Japanese) · Correct

t6, trial 1

Full text file (Japanese) · Incorrect

t6, trial 2

Full text file (Japanese) · Correct
Gemini

8 answers7 of them correct

t1, trial 1

Full text file (Japanese) · Incorrect

t2, trial 1

Full text file (Japanese) · Correct

t3, trial 1

Full text file (Japanese) · Correct

t3, trial 2

Full text file (Japanese) · Correct

t4, trial 1

Full text file (Japanese) · Correct

t4, trial 3

Full text file (Japanese) · Correct

t5, trial 1

Full text file (Japanese) · Correct

t6, trial 1

Full text file (Japanese) · Correct
Copilot

18 answers0 of them correct

t1, trial 1

Full text file (Japanese) · Incorrect

t1, trial 2

Full text file (Japanese) · Incorrect

t1, trial 3

Full text file (Japanese) · Incorrect

t2, trial 1

Full text file (Japanese) · Incorrect

t2, trial 2

Full text file (Japanese) · Incorrect

t2, trial 3

Full text file (Japanese) · Incorrect

t3, trial 1

Full text file (Japanese) · Incorrect

t3, trial 2

Full text file (Japanese) · Incorrect

t3, trial 3

Full text file (Japanese) · Incorrect

t4, trial 1

Full text file (Japanese) · Incorrect

t4, trial 2

Full text file (Japanese) · Incorrect

t4, trial 3

Full text file (Japanese) · Incorrect

t5, trial 1

Full text file (Japanese) · Incorrect

t5, trial 2

Full text file (Japanese) · Incorrect

t5, trial 3

Full text file (Japanese) · Incorrect

t6, trial 1

Full text file (Japanese) · Incorrect

t6, trial 2

Full text file (Japanese) · Incorrect

t6, trial 3

Full text file (Japanese) · Incorrect

MATERIALS

Materials used in the tasks

Original images we created ourselves: a receipt photo, a bar chart image, and a slightly tilted table image. The original numbers were set when we created them We recreated them for this run on November 2, 2026. Download them to repeat the same test yourself.

chart.png 24KB (Japanese)
receipt.png 67KB (Japanese)
table_photo.png 128KB (Japanese)

METHOD

How we measure

We list this run's conditions exactly as they were, so anyone can repeat it.

AIs measured and their model names at the time

AI services change under the hood without notice. We record the model name shown on screen and the time we checked it.

AIs measured and model names
AI	Plan	Model name shown on screen	Time checked
ChatGPT	ChatGPT Pro	GPT-6.1 Sol Medium	2026-11-02 21:03
Gemini	Google AI Pro	Flash	2026-11-02 21:03
Claude	Claude Max	Opus 4.8 High	2026-11-02 21:03
Copilot	Microsoft 365 Personal		2026-11-02 21:03

What we did in this run

We recreated the materials for the tasks specifically for this run. The data are for a fictional company and include no real companies or people.
We created 6 tasks and set the answer keys at that point.
We sent the same prompt and the same materials to ChatGPT, Gemini, Claude and Copilot.
We sent each task 3 times, in a new chat each time.
We rotated the sending order from run to run. This keeps the time of day from skewing the ranking.
We waited up to 300 seconds for each answer, and recorded anything longer as a timeout.
Scoring: Machine scoring only (not graded by AI).
We sent 72 trials in total, and 72 of them were valid.

How we count trials with no answer

Refused or timed out
We keep it in the base and count it as wrong. Because at work, that is the same as getting no answer.
Hit the plan's usage limit
We remove it from the base. We treat that trial as not measured.
A human verification screen appeared
We stop there. We neither click through it automatically nor try to get around it.
Operating error on AInformation's side
We remove it from the base. We do not report it as the AI's fault.

LIMITS

Limits of this benchmark

So that no one judges by the numbers alone, we list what we could not measure.

This is a single run on a single day. Rankings can change with the day or the task
The plan used is each AI's paid tier (ChatGPT Pro / Google AI Pro / Claude Max / Microsoft 365 Personal). These are not results from the free tier
Measurements use only the default model and default mode (features such as Deep Research are not used)
The tasks use original data for the fictional company Sample Shoji Co., Ltd. These are not results using data from a real company
Only tasks that can be scored by machine are covered. Writing quality and readability are not measured
Measurements are taken from the same screen a person would use in a browser. These are not results from the API
For any AI whose on-screen model name could not be read in a given run, that field is left blank (null). When the name could be read, it is shown as is

DATA

Published data

In the CSV, 1 row is 1 trial. The figures on this page match a recount of that CSV. The license is CC BY 4.0.

Latest version (CSV) (Japanese) 1 row = 1 trial, 16KB
Latest version (JSON) 38KB
v1 (version at publication) (CSV) (Japanese) 1 row = 1 trial, 16KB
v1 (version at publication) (JSON) 38KB

The cumulative rankings are at /en/data/research/bench/leaderboard.json. We do not delete versions (such as v1). That way, links from people who cited them keep working.

FAQ

How do you measure?: We send exactly the same task prompt and the same materials to the 4 AIs on the same day. We open a new chat every time so earlier answers don't carry over. The sending order is rotated in each run. We also record the seconds from sending until the answer stops.
Why do you send the same task 3 times?: Because AI answers vary even for the same question. From a single try, you can't tell whether a result reflects real ability or chance. We send each task 3 times, and the share of tasks whose conclusion matched is reported as the consistency rate.
Who does the scoring?: Scoring is done only by a program. We do not let AI score the answers. The correct answers are set when the tasks are created, and whether an answer is right is decided by whether numbers match, whether required words are present, whether the format fits, whether URLs actually open and the result of running formulas. Matters of taste are not scored.
Is the data in the tasks from real companies?: No. We create new data for a fictional company every time. We do not use real companies, real people or text from other companies. We share the materials we used as they are, so you can repeat the same test.
How do you count cases where the AI couldn't answer?: Refusals and timeouts stay in the base and are counted as wrong. If we hit the plan's usage limit or a human verification screen appears, we remove that trial from the base and treat it as not measured. Operating errors on AInformation's side are also removed from the base, and we do not report them as the AI's fault.
Do you use the free or paid versions?: We use each AI's paid plan through the browser interface. We don't use the API. For each run, we record the plan name, the model name shown on screen and the time we checked it.
Does this ranking show which AI is the smartest?: No. We measure only 5 things on a fixed set of work tasks: accuracy, time to answer, consistency rate, real-source rate and instruction compliance. We do not measure writing style, readability or behavior in long conversations. If the tasks change, the ranking changes too.
Can I cite these results?: Yes. The license is CC BY 4.0. As the source, please write “Source: AInformation, AI Practical Work Benchmark” with the URL of the page you used and the measurement date. The results are also available as JSON and CSV. With the CSV (1 row per trial), you can recount the figures we publish.

CITATION

Cite this benchmark

Source: AInformation, “AI Practical Work Benchmark: Reading numbers from images of receipts, charts, and tables (4 AIs, 6 tasks, 3 runs each)” (https://ainformation.jp/en/ai-surveys/ai-work-benchmark/numbers-from-images-2026-11-03). Measured on November 3, 2026. 6 tasks, 3 trials each, 72 valid trials

When you cite a figure, please include the base too (how many of how many trials). Please cite AInformation as the source, along with the URL of the page you used and the measurement date.

What this run found

Scores by AI

Results by task

Full answers from the 4 AIs

Materials used in the tasks

How we measure

AIs measured and their model names at the time

What we did in this run

How we count trials with no answer

Limits of this benchmark

Published data

FAQ

Cite this benchmark

Related pages