Skip to content

AI Model Benchmark

We sent exactly the same questions in Japanese to ChatGPT, Claude and Gemini and counted how many they answered correctly, how many seconds they took and how many tokens they used to answer. All scoring is automatic. These are not figures copied from official announcements.

Tested on September 5, 2026 / 3 trials per question / 8 models measured at the same time under the same conditions

This test was done in Japanese. The questions were sent to each company's API in Japanese, and the answers were scored in Japanese. The question text and scoring criteria below are AInformation's translations of the Japanese originals. Japanese original

8Models Number of real test runs
25Questions Questions scored by machine
600Trials Total number of trials
Filter
Provider
Model tier
Sort

Models shown: 8

Sorted by accuracy, highest first. We do not give an overall ranking, because which categories matter depends on what you use the AI for

GPT-6 Astra OpenAI Top tier
88% 66/75 Correct answers (out of 75 scored trials from 25 questions × 3 trials)
Instruction compliance 24/24
Factual accuracy 9/18
Calculation 18/18
Code 15/15
2.71sTime to finish
2.07sTime to first character
12,000Tokens used

Made-up answers: 0 When asked about things it didn't know, it said “I don't know” all 9 times

Claude Sonnet 5 Anthropic Everyday
88% 66/75 Correct answers (out of 75 scored trials from 25 questions × 3 trials)
Instruction compliance 24/24
Factual accuracy 9/18
Calculation 18/18
Code 15/15
2.22sTime to finish
1.58sTime to first character
33,653Tokens used

Wrong answers asserted: 1 When asked about things it didn't know, it made up numbers or details (times it said it didn't know: 8)

GPT-5.6 Sol OpenAI Top tier
88% 66/75 Correct answers (out of 75 scored trials from 25 questions × 3 trials)
Instruction compliance 24/24
Factual accuracy 9/18
Calculation 18/18
Code 15/15
2.35sTime to finish
1.69sTime to first character
17,321Tokens used

Wrong answers asserted: 3 When asked about things it didn't know, it made up numbers or details (times it said it didn't know: 6)

Gemini 3.1 Pro Google Top tier
88% 66/75 Correct answers (out of 75 scored trials from 25 questions × 3 trials)
Instruction compliance 24/24
Factual accuracy 9/18
Calculation 18/18
Code 15/15
6.98sTime to finish
6.98sTime to first character
95,683Tokens used

Made-up answers: 0 When asked about things it didn't know, it said “I don't know” all 9 times

Gemini 3.7 Flash Google Everyday
88% 66/75 Correct answers (out of 75 scored trials from 25 questions × 3 trials)
Instruction compliance 24/24
Factual accuracy 9/18
Calculation 18/18
Code 15/15
1.9sTime to finish
1.87sTime to first character
61,745Tokens used

Wrong answers asserted: 3 When asked about things it didn't know, it made up numbers or details (times it said it didn't know: 6)

GPT-5.6 Terra OpenAI Everyday
84% 63/75 Correct answers (out of 75 scored trials from 25 questions × 3 trials)
Instruction compliance 23/24
Factual accuracy 7/18
Calculation 18/18
Code 15/15
1.9sTime to finish
1.19sTime to first character
15,076Tokens used

Wrong answers asserted: 6 When asked about things it didn't know, it made up numbers or details (times it said it didn't know: 6)

Claude Opus 5 Anthropic Top tier
83.1% 59/71 Correct answers (out of 71 scored trials from 25 questions × 3 trials)
Instruction compliance 24/24
Factual accuracy 6/18
Calculation 18/18
Code 11/11
3.33sTime to finish
2.22sTime to first character
29,850Tokens used

Wrong answers asserted: 5 When asked about things it didn't know, it made up numbers or details (times it said it didn't know: 7)

Claude Fable 5.1 Anthropic Top tier
83.1% 54/65 Correct answers (out of 65 scored trials from 25 questions × 3 trials)
Instruction compliance 24/24
Factual accuracy 9/18
Calculation 16/18
Code 5/5
4.98sTime to finish
2.95sTime to first character
16,470Tokens used

Wrong answers asserted: 2 When asked about things it didn't know, it made up numbers or details (times it said it didn't know: 9)

Compare by category

These are the results of sending the same questions the same number of times. We do not give an overall score. Look only at the categories closest to the work you want to hand off. Bar length is the share of correct answers in that category.

Instruction compliance

Character count, line count, JSON format and banned words. Did it write exactly as instructed? / Models with every question right: 7

GPT-6 Astra 24/24
Claude Opus 5 24/24
Claude Fable 5.1 24/24
Claude Sonnet 5 24/24
GPT-5.6 Sol 24/24
Gemini 3.1 Pro 24/24
Gemini 3.7 Flash 24/24
GPT-5.6 Terra 23/24

Factual accuracy

Questions about prices and specs. We also check whether it said “I don't know” when it didn't know

GPT-6 Astra 9/18
Claude Fable 5.1 9/18
Claude Sonnet 5 9/18
GPT-5.6 Sol 9/18
Gemini 3.1 Pro 9/18
Gemini 3.7 Flash 9/18
GPT-5.6 Terra 7/18
Claude Opus 5 6/18

Calculation

Totals, tax, year-over-year changes and day counts. Is the number in the answer right? / Models with every question right: 7

GPT-6 Astra 18/18
Claude Opus 5 18/18
Claude Sonnet 5 18/18
GPT-5.6 Sol 18/18
GPT-5.6 Terra 18/18
Gemini 3.1 Pro 18/18
Gemini 3.7 Flash 18/18
Claude Fable 5.1 16/18

Code

We run the code it wrote and check whether the tests pass / Models with every question right: 8

GPT-6 Astra 15/15
Claude Opus 5 11/11
Claude Fable 5.1 5/5
Claude Sonnet 5 15/15
GPT-5.6 Sol 15/15
GPT-5.6 Terra 15/15
Gemini 3.1 Pro 15/15
Gemini 3.7 Flash 15/15

Models in lighter text had trials with no answer, so their base is smaller

All the numbers in tables

Same data as the cards above, for when you want to compare without re-sorting. “Tokens used” is the measured token count returned by each company's API, including tokens used while thinking. If two models get the same number right, fewer tokens means less waste.

Top-tier models compared

ModelsCorrect answers Instruction complianceFactual accuracyCalculationCode Time to finishTime to first characterTokens used
GPT-6 Astra OpenAI 66/75 24/24 9/18 18/18 15/15 2.71s 2.07s 12,000
Claude Opus 5 Anthropic 59/71 24/24 6/18 18/18 11/11 3.33s 2.22s 29,850
Claude Fable 5.1 Anthropic 54/65 24/24 9/18 16/18 5/5 4.98s 2.95s 16,470
GPT-5.6 Sol OpenAI 66/75 24/24 9/18 18/18 15/15 2.35s 1.69s 17,321
Gemini 3.1 Pro Google 66/75 24/24 9/18 18/18 15/15 6.98s 6.98s 95,683

Everyday models compared

ModelsCorrect answers Instruction complianceFactual accuracyCalculationCode Time to finishTime to first characterTokens used
Claude Sonnet 5 Anthropic 66/75 24/24 9/18 18/18 15/15 2.22s 1.58s 33,653
GPT-5.6 Terra OpenAI 63/75 23/24 7/18 18/18 15/15 1.9s 1.19s 15,076
Gemini 3.7 Flash Google 66/75 24/24 9/18 18/18 15/15 1.9s 1.87s 61,745

How models answer when they don't know

Results for questions about prices and about models that don't exist. Models that assert numbers they don't actually know can cost you if you take their answers at face value.

ModelsCorrect Said “I don't know”Asserted a wrong answer
GPT-6 Astra 9/18 9 0
Claude Opus 5 6/18 7 5
Claude Fable 5.1 9/18 9 2
Claude Sonnet 5 9/18 8 1
GPT-5.6 Sol 9/18 6 3
GPT-5.6 Terra 7/18 6 6
Gemini 3.1 Pro 9/18 9 0
Gemini 3.7 Flash 9/18 6 3

Which questions were missed (25 in total)

On 3 of the 25 questions, none of the 8 models got the answer right in all 3 trials. They all fall under Factual accuracy. For questions like these, plan on having a person check the answer.

For each question, we show how many of the 3 trials each model got right. The small text below shows the values we actually measured (how many characters the answer had, how many tests passed and so on). We do not show the answer text itself.

3/3All correct 1/3Varies by trial 0/3All wrong —No answer came back (not counted in the base)

Instruction compliance At least one model missed 1 of 8 questions
Question and scoring criteria GPT-6 AstraOpenAI Claude Opus 5Anthropic Claude Fable 5.1Anthropic Claude Sonnet 5Anthropic GPT-5.6 SolOpenAI GPT-5.6 TerraOpenAI Gemini 3.1 ProGoogle Gemini 3.7 FlashGoogle
Is it within ±5% of the specified 300 characters? follow-len-300 Correct answer: 285–315 characters (target: 300 characters) 3/33003/3308 / 3043/3304 / 3003/3303 / 3003/33002/3300 / 3013/3306 / 2993/3300 / 310
Whether the response is 100 characters or fewer follow-len-100 Correct answer: 100 characters or fewer 3/357 / 593/371 / 643/360 / 513/359 / 673/345 / 533/327 / 353/390 / 943/370 / 88
Whether it can be read as JSON and contains the specified keys and values follow-json Correct answer: All keys present and values match 3/3Match3/3Match3/3Match3/3Match3/3Match3/3Match3/3Match3/3Match
Whether there are exactly 5 bullet-point lines follow-bullet5 Correct answer: 5 lines 3/35 lines3/35 lines3/35 lines3/35 lines3/35 lines3/35 lines3/35 lines3/35 lines
Whether none of the banned words are used follow-ban Correct answer: No banned words used 3/3None3/3None3/3None3/3None3/3None3/3None3/3None3/3None
Whether the first 40 characters are free of any preamble follow-noprefix Correct answer: No preamble in the first 40 characters 3/3None3/3None3/3None3/3None3/3None3/3None3/3None3/3None
Whether the table has 3 columns and 3 rows follow-table Correct answer: 3 columns, 3 rows 3/33 columns, 3 rows3/33 columns, 3 rows3/33 columns, 3 rows3/33 columns, 3 rows3/33 columns, 3 rows3/33 columns, 3 rows3/33 columns, 3 rows3/33 columns, 3 rows
Whether the items are arranged in the specified order follow-order Correct answer: Gather requirements → Submit proposal → Create quote → Sign contract → Deliver 3/3In order3/3In order3/3In order3/3In order3/3In order3/3In order3/3In order3/3In order
Factual accuracy At least one model missed 5 of 6 questions
Question and scoring criteria GPT-6 AstraOpenAI Claude Opus 5Anthropic Claude Fable 5.1Anthropic Claude Sonnet 5Anthropic GPT-5.6 SolOpenAI GPT-5.6 TerraOpenAI Gemini 3.1 ProGoogle Gemini 3.7 FlashGoogle
Whether the numeric answer matches the correct value fact-price-own Correct answer: 12 0/3Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown0/315 / Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown0/30.3
Whether the numeric answer matches the correct value fact-price-rival Correct answer: 25 0/3Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown0/3Answered that it was unknown
Whether the numeric answer matches the correct value fact-price-jp Correct answer: 1400 0/3Answered that it was unknown0/3Answered that it was unknown / 11800/3Answered that it was unknown0/3Answered that it was unknown0/315000/31000 / 15000/3Answered that it was unknown0/3Answered that it was unknown
Whether yes/no matches the correct answer fact-free-plan Correct answer: Yes 3/3Yes3/3Yes3/3Yes3/3Yes3/3Yes1/3Yes / None3/3Yes3/3Yes
Whether it answers that the model does not exist, and does not fabricate details fact-nocheat Correct answer: States it doesn't exist and doesn't write its contents 3/3Said it didn't exist0/3Said it didn't exist but still wrote its contents (1 instance)3/3Said it didn't exist3/3Said it didn't exist3/3Said it didn't exist3/3Said it didn't exist3/3Said it didn't exist3/3Said it didn't exist
Whether the numeric answer matches the correct value fact-unit Correct answer: 3.3058 3/33.30583/33.30583/33.30583/33.30583/33.30583/33.30583/33.30583/33.3058
Calculation At least one model missed 1 of 6 questions
Question and scoring criteria GPT-6 AstraOpenAI Claude Opus 5Anthropic Claude Fable 5.1Anthropic Claude Sonnet 5Anthropic GPT-5.6 SolOpenAI GPT-5.6 TerraOpenAI Gemini 3.1 ProGoogle Gemini 3.7 FlashGoogle
Whether the numeric answer matches the correct value calc-sum Correct answer: 70530000 3/3705300003/3705300003/3705300003/3705300003/3705300003/3705300003/3705300003/370530000
Whether the numeric answer matches the correct value calc-tax Correct answer: 54480 3/3544803/3544803/3544803/3544803/3544803/3544803/3544803/354480
Whether the correct location name is included calc-pct Correct answer: Tokyo 3/3Tokyo3/3Tokyo3/3Tokyo3/3Tokyo3/3Tokyo3/3Tokyo3/3Tokyo3/3Tokyo
Whether the items are arranged in the specified order calc-share Correct answer: New installation → Regular maintenance → Consulting 3/3In order3/3In order3/3In order3/3In order3/3In order3/3In order3/3In order3/3In order
Whether the numeric answer matches the correct value calc-date Correct answer: 8 3/383/383/383/383/383/383/383/38
Whether the numeric answer matches the correct value calc-trap Correct answer: 49000000 3/3490000003/3490000001/3Couldn't identify the answer because there are 13 numbers / 15600.0, 15600000.0, 12400000.0, 9800000.0,3/3490000003/3490000003/3490000003/3490000003/349000000
Code At least one model missed 0 of 5 questions
Question and scoring criteria GPT-6 AstraOpenAI Claude Opus 5Anthropic Claude Fable 5.1Anthropic Claude Sonnet 5Anthropic GPT-5.6 SolOpenAI GPT-5.6 TerraOpenAI Gemini 3.1 ProGoogle Gemini 3.7 FlashGoogle
Whether running the written code actually passes all 6 tests code-zenkaku Correct answer: 6/6 3/36/6——3/36/63/36/63/36/63/36/63/36/6
Whether running the written code actually passes all 8 tests code-wareki Correct answer: 8/8 3/38/83/38/83/38/83/38/83/38/83/38/83/38/83/38/8
Whether running the written code actually passes all 6 tests code-yubin Correct answer: 6/6 3/36/62/26/6—3/36/63/36/63/36/63/36/63/36/6
Whether running the written code actually passes all 6 tests code-kingaku Correct answer: 6/6 3/36/63/36/62/26/63/36/63/36/63/36/63/36/63/36/6
Whether running the written code actually passes all 6 tests code-edge Correct answer: 6/6 3/36/63/36/6—3/36/63/36/63/36/63/36/63/36/6

How we measured

We submit the same Japanese questions without changing a single character. We don't set reasoning strength or temperature for any provider (each company's default is used). All tools such as web search are turned off.

Test date
September 5, 2026
Questions
25 questions in Japanese (Instruction compliance, Factual accuracy, Calculation and Code). The full text is below
Trials
3 trials per question. Times are medians
Output limit
16,000 tokens (same for all 3 companies)
Tool
Web search and other tools: None. We look only at the plain answers
Thinking effort
Each company's default setting (we did not set it for any company)
Where we ran it
From within Japan, directly to each company's API
Tokens used
The actual token count returned by each company's API. Includes tokens used for thinking
Scoring
All scored automatically.No subjective judgment by people or AI
Trials with no answer
We don't score them as 0; we leave them out of the base. That is why the base differs from model to model

Questions used (25 in total)

There's no way to check a benchmark whose contents you can't see. We show the exact text we sent and the scoring criteria as they are.

Instruction compliance8 questions

follow-len-300

Please write the body text of a “Notice of price revision due to rising raw material costs” letter to a business partner.
Condition: Write it in **exactly 300 characters**. Return only the body text, with no preamble or explanation.

Scoring: Is it within ±5% of the specified 300 characters?

follow-len-100

Write an internal notice saying that next week's all-hands meeting has been moved to Wednesday at 10 a.m. in Conference Room A.
Condition: **100 characters or fewer**. Return only the body text.

Scoring: Whether the response is 100 characters or fewer

follow-json

Extract the company name, amount, and payment due date from the following text.

"We have received an invoice for ¥1,254,000 for system maintenance fees from Midori Trading Co., Ltd. Payment is due by October 31, 2026."

Condition: Return only JSON. Use exactly three keys: company / amount / due. amount must be digits only (no commas). due must be in YYYY-MM-DD format. Do not add code block markers or any explanatory text.

Scoring: Whether it can be read as JSON and contains the specified keys and values

follow-bullet5

Summarize the following meeting notes into key points.

“The new system launches in October. First, only the accounting department will trial it. Since frontline staff said the input screen is hard to understand, it will be fixed by next month. Training will be held twice during September. The old system will be shut down at the end of December. The information systems department will handle inquiries.”

Condition: Use **exactly 5** bullet points. Each line must start with “- ”. Do not add an introduction or closing sentence.

Scoring: Whether there are exactly 5 bullet-point lines

follow-ban

Write a welcome greeting for new employees, around 200 characters.
Condition: **Never use the 3 words "ganbaru," "shikkari," or "zehi" even once.** Return only the body text.

Scoring: Whether none of the banned words are used

follow-noprefix

Write the body of an email to a business partner explaining that “because an item is out of stock, delivery will be delayed by about two weeks.”
Condition: **Do not write any greeting or preamble at all — return only the email body.** Do not write a reply to us either, such as “Understood” or “As follows.”

Scoring: Whether the first 40 characters are free of any preamble

follow-table

Choose three pieces of office equipment for internal use and create a comparison table.
Condition: Return **only a Markdown table**. Use exactly three columns: "Item," "Price range," and "Use." Use **exactly 3 rows**, not counting the header row. Do not add any text before or after the table.

Scoring: Whether the table has 3 columns and 3 rows

follow-order

Rearrange the following 5 tasks and write them out in the specified order.

Tasks: creating a quote / requirements interview / signing the contract / submitting a proposal / delivery

Condition: Arrange them in this order: **interview → proposal → quote → contract → delivery**. Write one task per line, task name only, for 5 lines. Do not add numbers or symbols.

Scoring: Whether the items are arranged in the specified order

Factual accuracy6 questions

fact-price-own

I'm asking about "GPT-6 Astra," the model I'm currently talking to.
When using this model via the API, what is the price per 1 million **output** tokens, in dollars?
Write only the number, in half-width characters, on one line (example: 25).
If you don't know, write only "unknown."

Scoring: Whether the numeric answer matches the correct value

fact-price-rival

When using OpenAI's "Claude Opus 5" via the API, what is the price per 1 million **output** tokens, in dollars?
Write only the number, in half-width characters, on one line.
If you don't know, write only "unknown."

Scoring: Whether the numeric answer matches the correct value

fact-price-jp

How much does ChatGPT's paid plan "Go" cost per month, in Japanese yen?
Write only the number, in half-width characters, on one line (example: 3000).
If you don't know, write only "unknown."

Scoring: Whether the numeric answer matches the correct value

fact-free-plan

Does "Gemini (Google AI plan)" have a plan that can be used for free forever (not a free trial, but a free plan)?
Write only one word, either "yes" or "no."
If you don't know, write only "unknown."

Scoring: Whether yes/no matches the correct answer

fact-nocheat

Tell me the context length and pricing for a model called "Claude Nimbus 7."
If it does not actually exist, just briefly say so.

Scoring: Whether it answers that the model does not exist, and does not fabricate details

fact-unit

How many square meters is 1 tsubo? Write only the number, in half-width characters, on one line, to 4 decimal places.

Scoring: Whether the numeric answer matches the correct value

Calculation6 questions

calc-sum

The table below shows a branch's monthly sales.

| Month | Sales (yen) |
|---|---|
| Jan | 4,820,000 |
| Feb | 5,140,000 |
| Mar | 4,390,000 |
| Apr | 6,020,000 |
| May | 5,570,000 |
| Jun | 5,930,000 |
| Jul | 6,480,000 |
| Aug | 4,960,000 |
| Sep | 5,310,000 |
| Oct | 6,750,000 |
| Nov | 7,120,000 |
| Dec | 8,040,000 |

What is the total sales for the year?
Write only the number, in half-width characters with no commas, on one line.

Scoring: Whether the numeric answer matches the correct value

calc-tax

The table below shows pre-tax amounts and the sales tax rate for each item.

| Item | Amount excl. tax (yen) | Tax rate |
|---|---|---|
| Meeting boxed lunches | 12,800 | 8% |
| Copy paper | 5,400 | 10% |
| Tea for guests | 3,200 | 8% |
| Office chair | 24,000 | 10% |
| Gift sweets | 4,500 | 8% |

What is the total of the tax-included amounts for all items?
Round down any fraction of a yen.
Write only the number, in half-width characters with no commas, on one line.

Scoring: Whether the numeric answer matches the correct value

calc-pct

The table below shows each location's sales for last year and this year.

| Location | Last year (yen) | This year (yen) |
|---|---|---|
| Tokyo | 128,400,000 | 141,240,000 |
| Osaka | 96,300,000 | 91,485,000 |
| Nagoya | 54,800,000 | 58,400,000 |
| Fukuoka | 38,200,000 | 37,054,000 |

Which location has the largest rate of change from last year?
Write only the location name on one line.

Scoring: Whether the correct location name is included

calc-share

The table below shows annual sales by service.

| Service | Sales (yen) |
|---|---|
| Regular maintenance | 41,300,000 |
| New installation | 68,900,000 |
| Maintenance parts | 22,400,000 |
| Training | 15,800,000 |
| Consulting | 33,600,000 |

List the top 3 services by sales, in descending order.
Write one service name per line, in descending order, for 3 lines.

Scoring: Whether the items are arranged in the specified order

calc-date

Between September 1, 2026 and September 30, 2026, how many Saturdays and Sundays are there?
Do not count public holidays. Give the total number of Saturdays and Sundays.
Write only the number, in half-width digits, on one line.

Scoring: Whether the numeric answer matches the correct value

calc-trap

The table below shows sales by branch. **Pay attention to the unit column.**

| Branch | Sales | Unit |
|---|---|---|
| Branch A | 12,400,000 | yen |
| Branch B | 9,800,000 | yen |
| Branch C | 15,600 | thousand yen |
| Branch D | 11,200,000 | yen |

What is the total sales for all 4 branches, in yen?
Write only the number, in half-width characters with no commas, on one line.

Scoring: Whether the numeric answer matches the correct value

Code5 questions

code-zenkaku

Write a Python function `to_hankaku(s)`.
It should convert full-width digits and full-width alphabetic characters to half-width, and return the result. Leave all other characters unchanged.
Return only the code.

Scoring: Whether running the written code actually passes all 6 tests

code-wareki

Write a Python function `to_seireki(s)`.
It takes a Japanese era-calendar string and returns the Western calendar year as an integer.
Support the Showa, Heisei, and Reiwa eras. Treat “gannen” (the first year of an era) as year 1.
Return only the code.

Scoring: Whether running the written code actually passes all 8 tests

code-yubin

Write a Python function `norm_zip(s)`.
It should normalize variant postal code formats into the form `123-4567` and return the result.
The input may mix full-width digits, full-width hyphens, a leading postal mark, and spaces.
If 7 digits cannot be extracted, return an empty string.
Return only the code.

Scoring: Whether running the written code actually passes all 6 tests

code-kingaku

Write a Python function `to_amount(s)`.
It takes an amount as a string and returns it as an integer. The string may contain commas, the yen kanji character, full-width and half-width yen symbols, and full-width digits. If there are no digits at all, return 0.
Return only the code.

Scoring: Whether running the written code actually passes all 6 tests

code-edge

Write a Python function `sum_safe(xs)`.
It should return the sum of a list of numbers.
It must not crash and must still return a sum when the list is None, when it is empty, or when it contains None or empty strings mixed in. Treat any element that cannot be converted to a number as 0.
Return only the code.

Scoring: Whether running the written code actually passes all 6 tests

Past tests

We keep a record each time we measure again.Each test covered different models and a different number of trials, so the results are not a comparison on the same scale. Read them together with the conditions at the time. Tests in which no answer came back at all are not listed.

September 5, 2026 13:25 8 models / 25 questions / 3 trials per question

Gemini's API key started working, so we re-measured 8 models from all three companies

GPT-6 Astra 88%
Claude Sonnet 5 88%
GPT-5.6 Sol 88%
Gemini 3.1 Pro 88%
Gemini 3.7 Flash 88%
GPT-5.6 Terra 84%
Claude Opus 5 83.1%
Claude Fable 5.1 83.1%
September 5, 2026 08:28 6 models / 25 questions / 3 trials per question

GPT-6 Astra became available (Sep 5).

GPT-6 Astra 88%
GPT-5.6 Sol 88%
GPT-5.6 Terra 88%
Claude Opus 5 87.1%
Claude Sonnet 5 84%
Claude Fable 5.1 83.3%
September 5, 2026 08:19 5 models / 25 questions / 3 trials per question

GPT-6 Astra became available via API (Sep 5), so we re-measured with 7 models

GPT-6 Astra 88%
GPT-5.6 Sol 88%
Claude Opus 5 87.5%
Claude Sonnet 5 86.7%
GPT-5.6 Terra 86.7%
September 4, 2026 08:15 4 models / 25 questions / 1 trials per question

Actual measurement for the GPT-6 Astra article (gap between the announcement and reality)

GPT-5.6 Sol 88%
GPT-5.6 Terra 88%
Claude Fable 5.1 86.4%
Claude Opus 5 83.3%
September 4, 2026 08:07 1 models / 25 questions / 1 trials per question

A check to confirm the scoring is reasonable

Claude Sonnet 5 84%

Official pages

The numbers above arewhat we counted ourselves by actually sending the questions. For the specs each company publishes, see the links below.

ModelsDate checkedSource
GPT-6 Astra September 5, 2026 Official page
Claude Opus 5 September 4, 2026 Official page
Claude Fable 5.1 September 2, 2026 Official page
Claude Sonnet 5 September 2, 2026 Official page
GPT-5.6 Sol September 2, 2026 Official page
GPT-5.6 Terra September 2, 2026 Official page
Gemini 3.1 Pro September 2, 2026 Official page
Gemini 3.7 Flash September 2, 2026 Official page

FAQ

Where do these numbers come from?

AInformation counted them on September 5, 2026 by actually sending the same questions to each company's API. They are not copied from other benchmarks or from the companies' announcements. We sent 25 questions to 8 models, 3 times each, for 600 trials in total.

Why is there no overall ranking?

A ranking means deciding how much weight to give each category, and that differs from person to person. Someone who needs the AI to follow instructions exactly and someone who wants it to handle calculations should look at different categories. So we show only the numbers for each category and leave the judgment to you.

How is scoring done?

We use only questions that can be scored entirely by machine: whether the character count matches the instruction, whether the output can be read as JSON, whether the calculated answer is right and whether the code it wrote actually runs. We do not include subjective criteria such as whether the writing is good or sounds natural. All scoring criteria are shown on this page.

What does “Tokens used” mean?

It is the actual token count returned by each company's API. A token is the unit an AI uses to handle text; in Japanese, 1 character is roughly 1 token. If two models get the same number of questions right, the one that used fewer tokens did less unnecessary thinking. Tokens used while thinking are included in the count.

When is it updated?

We measure again when a new model comes out. There is no fixed schedule. When we do, we remeasure every model on the page, not just the ones from 1 company. Numbers measured on different days would not be a comparison under the same conditions.

Are the questions public?

The full text of every question and the scoring criteria are at the bottom of the page, because a benchmark you can't see inside can't be checked.

About the data on this page

AInformation researched the number of correct answers, the seconds until each answer was finished and the tokens used, measured by giving major AI models the same questions in Japanese as of September 5, 2026. This data covers 8 models.

What AInformation found: We sent the same 25 questions to 8 models, 600 times each, and measured the results.

What this page shows
the number of correct answers, the seconds until each answer was finished and the tokens used, measured by giving major AI models the same questions in Japanese
Number of records
8 models
Research date
September 5, 2026 (last updated)
Survey method
We send exactly the same prompts to each company's API and judge the replies automatically. No subjective judgment by people or AI is involved. Output limits, tool use and reasoning effort are set the same for every company. The full text of the prompts and the scoring criteria are on the page.
What we collect
AInformation (site operator: Shunta Fujii). About us and our editorial policy
Machine-readable data
https://ainformation.jp/en/data/ai-benchmark.json (JSON, CC BY 4.0. How to use it)
How to cite

Source: AInformation, “AI Model Benchmark” (https://ainformation.jp/en/ai-model-benchmark), as of September 5, 2026

NEWSLETTER

The latest AI news, rounded up once a week

One email with the week's main AI news and the tools AInformation's editors actually tried.