Tokens: Astra 2.58M, Sol 2.67M, Opus 4.56M, Fable 7.17M

Tokens in the logs, child tasks included. The dark tick is the figure the model itself reported last

OpenAIAstra
2.58M, reported about 1.42M
OpenAISol
2.67M, reported about 2M
AnthropicOpus
4.56M, reported 3.9M
AnthropicFable
7.17M, reported about 6.9M
← Library · Model comparisons Model comparison · web search

GPT-6.2 Astra vs GPT-6.2 Sol vs Opus 5.5 vs Fable 5.2Researching workplace and business-partner relationships in Japan

We compared GPT-6.2 Astra and GPT-6.2 Sol from OpenAI with Claude Opus 5.5 and Claude Fable 5.2 from Anthropic. At 10:40 JST on 18 November 2026, we sent the same research request, written in Japanese, to all four models within 40 seconds. Every model ran at the "high" reasoning setting. All four used the same research skill (d-search) with the instruction to keep costs down (go cheaper).

There is one conclusion. When you hit a wall, try another model and see whether a different direction comes out.

18 Nov 2026

Video · 6:02
Our takeaways
  • When you hit a wall, try another model and see whether a different direction comes out.
  • This time, the clear difference was where each model went looking. Three reached the same conclusion, and only Astra went looking for how the staff on the two sides of a business relationship behave toward each other. Adding one more model can bring out a different direction like this.
  • Honestly, there is no best model. Even pinning down what each one is like is hard. Most of the differences came from the conditions of each run. The token gap came from what each run used for its child tasks, and Fable's larger number of facts came from its extra research.
  • Naming the direction you want searched in the request is another step you can take before trying a different model.
  • Whatever the model, you still need to check the sources. In 3 of the 30 cited passages, the figures the model wrote under the quote did not match.
Brief · what to pick

Which run for which job.

Work time runs across, and tokens in the logs run up. A bigger point means more facts in the research note. Toward the lower left is faster and uses fewer tokens.

Facts in the note102030Fable up to its first report (12 min 35 s, 4.58M tokens, 18 facts)
OpenAIAstra7 min 30 s, 2.58M tokens, 10 facts in the noteOpenAISol8 min 24 s, 2.67M tokens, 7 facts in the noteAnthropicOpus9 min 13 s (not counting 9 s waiting for our answer to its question), 4.56M tokens, 13 facts in the noteAnthropicFable18 min 20 s (not counting 11 min 30 s waiting for our answer), 7.17M tokens, 27 facts in the note

Each model ran once.

Prompt

The prompt sent to all four runs.

As sent, in Japanese

日本の職場で発生する人間関係の問題について調査して
取引先との人間関係の問題も含む

ファクトとAI推論は分けて出力
アウトプットは日本語

$d-search
go cheaper

English translation

Research the interpersonal problems that occur in Japanese workplaces
Include interpersonal problems with business partners

Output facts and AI inferences separately
Output in Japanese

$d-search
go cheaper

"$d-search" calls Dorado's research skill. It quotes the original text of each source and keeps confirmed facts apart from the AI's inferences. "go cheaper" asks for cheaper research methods without changing the scope or depth.

Measured · elapsed time

Which model finished first?

7 min 30 sAstraanswered at 10:48
8 min 24 sSolanswered at 10:48
9 min 13 sOpusanswered at 10:49
18 min 20 sFableanswered at 11:09
Work time

Each bar is the work time from sending the prompt to the last answer. Time the model spent waiting for our reply is not in the bar and shows as a dotted line. In Fable's bar, the dark part runs up to its first report and the light part is its extra research.

OpenAIAstra

7 min 30 s. Answered at 10:48

OpenAISol

8 min 24 s. Answered at 10:48

AnthropicOpus
waiting 9 s

9 min 13 s. Answered at 10:49. Right after starting, it asked, "The same request is running in two other threads. Is this for a comparison?" It got our reply 9 s later and carried on

AnthropicFable
To first report 12 min 35 s
Extra research 5 min 45 s
waiting 11 min 30 s

18 min 20 s. First report at 10:52 (18 facts). It had reached its 4M-token budget, so it asked whether to research 5 more items. It got our reply, "research the 5 items", 11 min 30 s later and sent its final report at 11:09

0 min5 min10 min15 min20 min25 min30 min
Fable: up to its first reportFable: extra researchTime spent waiting for our reply

The two GPT-6.2 models asked nothing and ran straight through to the end. The two Claude models each asked one question along the way. Opus asked to check for duplicates, and Fable asked about its budget. Counting the 11 min 30 s until we replied, Fable's final report came 29 min 50 s after we sent the prompt.

Measured · tokens

How many tokens did each run use?

2.58MAstrareported about 1.42M
2.67MSolreported about 2M
4.56MOpusreported 3.9M
7.17MFablereported about 6.9M
Tokens in the main session and the child tasks

Usage in the logs. The dark part is the main session, and the light part is the child tasks (helpers). The black tick is the figure the model itself reported last.

OpenAIAstra
Main 1.76M
0.81M

2.58M (main 1.76M, 0.81M in a child task on Sol low). Reported about 1.42M

OpenAISol
Main 2.29M

2.67M (main 2.29M, 0.38M in a child task on Sol low). Reported about 2M

AnthropicOpus
Main 2.75M
1.8M

4.56M (main 2.75M, 1.8M in a child task on Sonnet 5.5). Reported 3.9M

AnthropicFable
Main 2.69M

7.17M (main 2.69M, 4.48M in 4 child tasks on Sonnet 5.5). Reported about 6.9M

02M4M6M8M
Main sessionChild tasks (Astra, Sol: GPT-6.2 Sol low. Opus, Fable: Sonnet 5.5)The model's own report

All four reported less than the logs show. The gap was largest for Astra, which reported 55% of the logged total. Astra itself wrote that "the delegate's usage is not included". Sol reported 75% of its logged total, Opus 86%, and Fable 96%.

Both GPT-6.2 models set a budget of 1.5M tokens before they started researching, and both used 1.7 to 1.8 times that. Opus used 4.56M against a 4M budget. Fable stopped once at 4M, and then used 7.17M against a budget of 5.5M with the extra 1.5M added.

The two GPT-6.2 models used GPT-6.2 Sol at the low setting for their child tasks. The two Claude models used Sonnet 5.5 for theirs.

Measured · link check of the answers

Does each chat answer hold up as written?

After the runs, we opened every cited link in the answers and compared each page with the passage that cites it. A passage counts as backed when every figure, name, date, and condition in it is on the page.

Of the 30 cited passages in the four answers, 27 were fully backed. The other 3 were partly backed. None were unbacked, and none were dead links.

7 of 7Astrafully backed
6 of 6Solfully backed
5 of 6Opusfully backed
9 of 11Fablefully backed
How well the cited passages are backed

For each model, the share of its cited passages that were fully backed, partly backed, or not backed or unverifiable.

OpenAIAstra
7

7 of 7

OpenAISol
6

6 of 6

AnthropicOpus
5
1

5 of 6

AnthropicFable
9
2

9 of 11

0%25%50%75%100%
BackedPartly backedNot backed or unverifiable

In all 3 partly backed passages, the error was not in the quote itself. It was in the note the model added under the quote.

  • AnthropicFable: in its note on the number of workers' compensation awards, it wrote that "the top three were all about relationships with people". The third item in the quote is "a major change in the content or amount of work", which is not about relationships with people.
  • AnthropicFable: it gave the number of companies answering the company survey as "1,880". The figure in the chart is 1,870.
  • AnthropicOpus: it gave the respondents to the power harassment question as "n=7,774". 7,774 is the number of companies answering the company survey. The worker survey had 8,000 respondents.

The dates and source locations also had small errors.

  • Seven passages dated the survey report "17 May 2024" (Fable 3, Astra 2, Sol 2). The PDF cover is dated March 2024 (Reiwa 6), and 17 May appears nowhere in the PDF.
  • Fable wrote "no publication date given" for the Industrial Safety and Health Survey. Page 1 of the PDF gives 7 August 2026 (Reiwa 8).
  • Opus wrote that "the Ministry of Health, Labour and Welfare's own page still described the effective date in old wording" and cited a regional labour bureau's page instead. On the day we checked, the ministry's own page gave the effective date of 1 October 2026 (Reiwa 8).

The Ministry of Economy, Trade and Industry page returned a 403 error on the first automated fetch. It opened when we fetched it again with the same settings as a browser, so we count it as backed.

Compared · what each answer found

Where the four answers agree, and where they part.

Facts in the answers

The facts in the four answers, grouped as "inside the company", "business partners", and "terms of trade". A dot means that model wrote the fact in its answer.

Fact written in the answerOpenAIAstraOpenAISolAnthropicOpusAnthropicFableTotal
Inside the company
Used the Survey on Workplace Harassment (MHLW-commissioned, FY2023)4
32.9% named "interpersonal relationships" as a cause of strong stress (Industrial Safety and Health Survey, 2025)3
19.3% had experienced power harassment (same harassment survey)3
65.7% of power harassers were bosses2
36.9% "did nothing" after power harassment2
Mental-disorder compensation awards: power harassment by a boss most common, 222 cases (FY2025)2
"Workplace relationships" as a reason for leaving the previous job (Survey on Employment Trends, 2025)1
Top real reason for leaving is "bad relationships" (En Japan, 2024)1
Consulted someone about a relationship problem 20.3%, did not 14.6% (JILPT)1
27.7% have no one nearby to consult (Job Soken)1
Sexual harassment: sexual jokes 49.7%1
Business partners
22.6% named people at business partners as perpetrators of nuisance behavior2
Customer harassment measures an employer obligation from 1 October 20262
The MHLW counts a business partner's staff among the "customers and others" in customer harassment1
Nuisance behavior by customers 46.8% (UA Zensen, 2024)1
Mainly business-to-business dealings: 16 cases of verbal abuse or intimidation, 5 of being kept for long periods (Hamamatsu CCI)1
Cases of being unable to refuse a business partner's request (Food Rengo, UA Zensen)1
67.0% sent staff with future business in mind (JFTC, 2012)1
Terms of trade
Price negotiations 90.7% or pass-through rate 54.2% (SME Agency, March 2026)3
6.9% held back from price negotiations for fear of fewer orders (same survey)1
Act on Proper Trade: 8,261 guidance cases, 39 recommendations (JFTC, FY2025)1
Freelancers' fees reduced 28.1% (JFTC and MHLW, 2024)1

1 fact used by all four, 3 facts used by three, 5 facts used by two, and 13 facts used by only one (Astra 6, Fable 4, Sol 3, Opus 0).

The only source all four answers used was the "Survey on Workplace Harassment" (FY2023), commissioned by the Ministry of Health, Labour and Welfare. Three other facts appeared in three or more answers: 32.9% of workers named interpersonal relationships as a cause of strong stress, 19.3% had experienced power harassment, and the figures on price negotiation and passing on costs.

On relationships with business partners, the answers split in two.

  • Opus, Fable, and Sol looked for government statistics. They concluded that problems with business partners are counted in two separate groups: "nuisance behavior by people at a business partner (customer harassment)" and "terms of trade such as prices and fees". Opus wrote, "We did not find any government statistic that counts strained relationships with business partners themselves (things that are hard to say, awkwardness)."
  • Astra looked for sources that directly show how the staff on the two sides of a business relationship behave toward each other. It found verbal abuse, intimidation, and being kept for long periods in business-to-business dealings (a members' survey by the Hamamatsu Chamber of Commerce and Industry), a case where staff could not refuse a purchase request that cited last year's figures (a survey by Food Rengo and UA Zensen), and the share of suppliers that sent staff to a retailer with future business in mind (Japan Fair Trade Commission, 2012). For every one of these sources, Astra itself noted the limits of its scope in the table.

Sol picked out one figure from the trade-terms statistics that is close to relationships between people. 6.9% of answers said the company did not ask to negotiate prices because it feared fewer orders or the end of the business relationship.

13 facts appeared in only one answer: 6 from Astra, 4 from Fable, 3 from Sol, and none from Opus. Every fact in Opus's answer also appears in Fable's or Sol's.

The answers also differed in how they wrote their inferences.

  • AnthropicOpus: sorted out what the government statistics count and what they do not.
  • AnthropicFable: wrote 10 inferences. One example: "The side with more power may count what it does at a lower rate than the other side does."
  • OpenAISol: wrote only 2 inferences. Both are cautions on how to read in-house consultation counts and how business partners react.
  • OpenAIAstra: wrote them as what readers should look at in their own workplace. One example: "The fact that someone accepted a request does not show that they agreed with it."
The main result · research notes

The research note each run wrote.

The answer each run returned in chat is a summary of the research note it wrote. For each fact, the note holds the source's original text and where in the source it appears. You can read all four notes in full under "Actual results" below.

  • OpenAIAstra: 10 facts, 148 lines, 17,958 characters, 12 source links
  • OpenAISol: 7 facts, 99 lines, 12,691 characters, 7 source links
  • AnthropicOpus: 13 facts, 170 lines, 23,842 characters, 14 source links
  • AnthropicFable: 27 facts, 342 lines, 41,540 characters, 17 source links (no duplicates)
All four answers

All four answers, with their numbers.

ItemOpenAIGPT-6.2 Astra HighOpenAIOpenAIGPT-6.2 Sol HighOpenAIAnthropicClaude Opus 5.5 HighAnthropicAnthropicClaude Fable 5.2 HighAnthropic
Tokens used (logs)2,576,5002,670,2984,557,4287,172,440
Tokens the model reportedabout 1.42Mabout 2M3.9Mabout 6.9M
Token budget1.5M1.5M4M4M + 1.5M added
Work time7 min 30 s8 min 24 s9 min 13 s18 min 20 s
Questions along the wayNoneNoneOnce (duplicate check, waited 9 s)Once (budget check, waited 11 min 30 s)
Child tasks1 on GPT-6.2 Sol low1 on GPT-6.2 Sol low1 on Sonnet 5.54 on Sonnet 5.5
Answer length (Japanese characters)2,5382,3562,8526,181
Facts in the answer76611
Inferences in the answer32410
Cited passages checked76611
Fully backed7659
Partly backed0012
Not backed or dead link0000
Facts in the research note1071327
All four outputs

Actual results

Each run's research note and final answer are shown unedited. Switch between the models, and between "Research note" and "Final answer", to compare them. Fable's answer joins its three messages: the first report (sent in two parts) and the final report.

In Japanese workplaces, the reported problems include having no one to consult when needed, getting no advice that would help improve the work, and verbal abuse or exclusion by bosses and colleagues. With business partners, the research confirmed verbal abuse and being kept for long periods, and being unable to refuse a partner's request with future business in mind.

Facts

Relationship or situationWhat the surveys confirmed
Not consulting anyone when a problem arises at workOf 3,049 employees surveyed by the Japan Institute for Labour Policy and Training in 2018, those who had a problem with workplace relationships in the past year and consulted someone were 20.3%, and those who did not consult anyone were 14.6%. Both are shares of all 3,049 employees. 3rd Comprehensive Survey on the Working Lives of Japanese People, Figure 1-2-10-14, published 27 April 2026
No one nearby to consult, no adviceWhen Job Soken asked 578 working registered users about stress from relationships, 27.7% said "there is no one nearby I can consult", 26.8% said "the line on harassment is hard to judge", and 26.5% said "there is no constructive feedback". 2025 Workplace Stress Survey, 17 March 2025
Power harassment by bosses and colleaguesIn the Ministry of Health, Labour and Welfare survey, of 1,546 people who said they had experienced power harassment in the past three years, 65.7% named a boss other than an executive as the harasser, and 20.2% named a colleague. The behavior was threats, insults, severe verbal abuse, and the like in 48.5% of cases, excessive demands in 38.8%, and isolation, exclusion, or being ignored in 27.8%. All are multiple-answer questions. MHLW, harassers, Figure 149, 17 May 2024 / types of behavior, Figure 137, same day
Sexual conduct at workIn the same MHLW survey, of 503 people who said they had experienced sexual harassment in the past three years, 49.7% named "sexual jokes or teasing" and 26.2% named "unnecessary physical contact". Multiple answers. MHLW, Survey on Workplace Harassment, Figure 138, 17 May 2024
Verbal abuse, intimidation, and being kept for long periods in business-to-business dealingsIn a survey of member businesses by the Hamamatsu Chamber of Commerce and Industry, the group dealing mainly with corporate customers reported 16 cases of "verbal abuse or intimidation" and 5 of "being kept for long periods". These are counts of answers in a regional members' survey. Hamamatsu Chamber of Commerce and Industry, results of the questionnaire on customer harassment, 1 October 2026
Unable to refuse a business partner's purchase requestA survey of union members working in sales in food manufacturing and in distribution and retail includes this case: "A sales representative at a business partner cites last year's figures and asks me to report how many seasonal products we will buy, and I can't refuse." The text does not show whether the partner said outright that it would cut business if refused. Food Rengo and UA Zensen, "Survey on Trading Practices", result 10, February 2025
Sending staff with future business in mindIn a Japan Fair Trade Commission survey, of 581 suppliers that answered the question after large retailers and others asked them to send employees, 389 (67.0%) answered that "considering future business, there were times we had no choice but to agree to the request". It is a 2012 survey of retail transactions, and the current share across all industries is unknown. JFTC, Survey on Transactions between Large Retailers and Suppliers, Figure 35, 11 July 2012

AI inferences

  • Judging relationships by the number of in-house consultations alone will mislead you. Some people did not consult anyone even when a problem actually arose. Others named the lack of anyone nearby to consult as a source of stress in itself. However, these are two separate surveys, so they cannot be linked to explain why people did not consult anyone.
  • Narrowing in-house problems to "how the boss talks" misses being ignored by colleagues and how work is handed out. Besides who said what, look separately at whether the person was left out of places they needed to be and whether they were asked for work they could not do. That makes it easier to see what the person is struggling with.
  • In a business relationship, the fact that someone accepted a request does not show that they agreed with it. Asking how many seasonal products the other side will buy, or asking for extra staff, looks like normal business to the side making the request. For the side receiving it, whether it can keep trading on the same terms after refusing may decide the answer.

Full research report (original text, sources, limits of each survey)

The measured use of the parent thread was about 1.42M tokens, 19 steps, about 6 minutes. The budget was 1.5M tokens, and the delegate's usage is not included in the measurement.

The research note link at the end of each answer opens that note's download. Links to other chat threads keep only their text.

The answers are AI translations of the Japanese originals, which are the record. The research notes are shown in the original Japanese.

The research plan
$50 first month, then $200 a month

Want the sources checked for your product?

These answers are unchecked. In the research plan, we open the sources, check the dates, and correct what is wrong before the brief reaches you.