Fable + Luna used 16.1M tokens, Opus 11.7M, Sonnet 10.7M, Sol 8.4M, Fable 5.3M, Astra 2.9M

Tokens used, helpers included. The dark tick is the figure each model reported at the end.

OpenAIAstra
2.9M, reported 0.7M
AnthropicFable
5.3M, reported 4.6M
AnthropicOpus
11.7M, reported 9.8M
AnthropicSonnet
10.7M, reported 9.9M
OpenAISol
8.4M, reported 2.9M
AnthropicFable + Luna
16.1M, reported 15.4M
← Library · Model comparisons Model comparison · research loop

Fable 5.2 vs GPT-6 Astra vs Opus 5.5 vs Sonnet 5.5 vs GPT-6.1 SolJapan's growth industries, by CAGR

Claude Fable 5.2 Medium from Anthropic and GPT-6 Astra Medium from OpenAI got the same Japanese research request on the same morning. Claude Opus 5.5 High from Anthropic got it at midday, and Claude Sonnet 5.5 High from Anthropic and GPT-6.1 Sol High from OpenAI the next morning. All five used the same research skill with up to three rounds, and Sonnet alone added one more round with approval. A sixth run, later that morning, added execution conditions to the same request: Fable 5.2 Medium planned and wrote, and child tasks on GPT-6 Luna High from OpenAI did the searching and page reading. This page compares how long each run took, how many tokens it used, and whether its cited pages back what it wrote. The answers are in Japanese. This page shows them as AI translations, and the original is one click away.

14 Nov 2026

Our takeaways
  • Scan with Fable. Deep-dive with Opus, or with Sol on a budget.
  • Opus is too meticulous to do broad research.
  • Astra can still get a good overview.
  • Using the models in combination is still the best approach.
Brief · what to pick

Which run for which job.

Lower left is cheaper and faster.

Findings (sourced quotes) =1050100= Deep
Broad, fast → FableBroad, cheap → SonnetDeep, fast → OpusDeep, cheap → SolOverview only → AstraNot picked → Fable + Luna

Thick ring: recorded where every quote sits and took margins from earnings releases (Opus), or every cited passage backed (Sol). One run per combination. Fable's helpers were Sonnet. The reasons are in the sections below.

Prompt

The prompt used for all six runs.

As sent, in Japanese

では、2026年から向こう3年で伸びる日本の業界を調べてきてほしい

CAGRだね
大きな大局を調べつつも、ニッチでも伸びている、みんなが見てないところで儲かっている業界があるのでそういうのも調べてきて
$d-loop-search x 3 go cheaper

English translation

Next, I'd like you to research the Japanese industries that will grow over the next three years from 2026.

Use CAGR.
Cover the big picture, but there are also industries that are growing in niches and making money where nobody is looking, so research those too.
$d-loop-search x 3 go cheaper

Added after the request above for the sixth run, Fable + Luna, only

---
実行条件(モデル比較のテスト。上の依頼文は前回の5モデルと同じ):
- 各ラウンドの検索とページの読み込みは、delegate_task で Codex の gpt-6-luna(推論 high)の子タスクに1つずつ任せる。子タスクは3つまで。
- 本体(あなた)は、計画、ラウンドごとの穴の確認、最終の回答とノートの作成だけを行う。本体ではウェブ検索もページの取得もしない。子タスクの要約と、保存された引用だけを読む。
- 引用の照合は、本体が読むのではなく、スクリプトで保存ページと突き合わせる。
- research/ にある 2026-11-14 と 2026-11-15 の成長業界のノートと、dorado/content/model-comparisons/ は開かない(比較のため)。
- ノートは research/2026-11-15-japan-growth-fable-luna.md に保存する。最後に、使ったトークン(本体と子タスク別)と所要時間を報告する。

English translation of the conditions

---
Execution conditions (a model-comparison test. The request above is the same as for the earlier five models):
- Hand the searching and page reading of each round, one at a time, to a child task on Codex's gpt-6-luna (reasoning high) with delegate_task. Up to three child tasks.
- The main session (you) only plans, checks the gaps after each round, and writes the final answer and the note. The main session does no web search and fetches no pages. It reads only the child tasks' summaries and the saved quotes.
- Check quotes against the saved pages with a script, not by reading them in the main session.
- Do not open the growth-industry notes of 2026-11-14 and 2026-11-15 in research/, or dorado/content/model-comparisons/ (for the comparison).
- Save the note to research/2026-11-15-japan-growth-fable-luna.md. At the end, report the tokens used (main session and child tasks separately) and the elapsed time.

"$d-loop-search x 3" tells the model to run Dorado's research skill for up to three rounds, each round filling the gaps of the last. "go cheaper" asks it to keep token use down. Astra and Fable started at 10:15 on 14 November 2026. Each model planned the research itself, handed the searching to helpers on a cheaper model, checked its quotes against the saved pages, and wrote the answer in Japanese. Opus started at 12:05 the same day, with "/d-loop-search", the Claude Code form of the same command. Before researching, Opus found Fable's note on the same question, read its opening, and asked Minoru whether to research again. Minoru told it to start over. Sonnet started at 08:37 the next day, 15 November, with the same text. It did not open any of the three notes from the day before. After three rounds, before it started a fourth helper, Dorado's token-control hook asked Minoru for approval. Minoru answered "go cheaper", and Sonnet added only round 5, limited to niches and profit. Sol started at 08:39 the same morning with the same text and handed each of its three rounds to a T3 child task (two GPT-6 Luna High, one GPT-6.1 Sol Medium). Sol did not open the earlier notes either. The sixth run, Fable + Luna (Fable 5.2 Medium), started at 10:35 the same day with the same text followed by the execution conditions below. The conditions gave all searching and page reading to three child tasks on GPT-6 Luna High, limited the main session to planning, checking the gaps after each round, and writing the answer and the note, and had a script check the quotes instead of the main session reading the pages. The other five runs got no such conditions. The logs show no step that opened the notes from that day or the day before, or the comparison folder.

The main result · research notes

The research note each run wrote.

The answer each run returned in chat summarizes the research note it wrote. The note holds, for each finding, the quoted source text and its link, plus the coverage status, the source audit, and a record of each round. All six notes can be read in full under Actual results below.

CheckOpenAIGPT-6 Astra MediumOpenAIAnthropicClaude Fable 5.2 MediumAnthropicAnthropicClaude Opus 5.5 HighAnthropicAnthropicClaude Sonnet 5.5 HighAnthropicOpenAIGPT-6.1 Sol HighOpenAIAnthropicClaude Fable 5.2 Medium + GPT-6 Luna HighAnthropic · OpenAI
Note length280 lines1,413 lines1,223 lines910 lines403 lines913 lines
Findings (with quote and source)11 (plus 3 pilot records)987848 (4 of them overflow records beyond the 3-per-publisher cap)1651 (from 55 checked, after removing 4 duplicates and others)
Distinct links14154981091645
Websites cited6513744920
Where each quote sits (section path)RecordedNot recorded for anyRecorded for allPartly recorded. Not recorded for the 16 from round 1Recorded for allRecorded for 50 of 51
Coverage items6 items. States that no 2026 to 2029 ranking could be made6 items: 2 covered, 4 partly covered7 items: 3 covered, 4 partly covered8 items: 2 covered, 6 partly covered9 items: 2 covered, 2 covered for the adopted candidates, 5 partly covered8 items: 1 covered, 7 partly covered
The model's own quote checkChecked the 11 against saved originals and recorded all as backedA script matched all 98 quotes to the saved pagesA script matched each round's quotes to the saved pages. 13 table rows were matched by the order of their figuresA script matched the quotes to the saved pages as strings. Table rows were checked by reading the page againA script matched quotes to the saved originals, and the main session checked subject, period, and unit. Recorded all 23 linked lines as backedA script matched all 55 quotes to the pages the child tasks saved and recorded every one as matching. The main session did not read the pages
What the model's own audit fixedCorrected a mix-up of MonotaRO's standalone and consolidated figuresRecorded per round, for example F5 removed, G17 downgraded, and the growth rates of F4 and G5 removedRemoved 1 finding (R1-31) and dropped 1 figure that rested only on a search snippetWithdrew the calculated CAGR in F10, dropped a start value from a search summary in F3, and left F35 and F36 out of the conclusionsWithdrew 2 profit figures read only in search results, dropped Organo's profit, and corrected network security's CAGR from 8.03% to 5.93%Downgraded 8 to partly backed, removed 4, and replaced the publication dates the child tasks recorded with the dates on the saved pages
Dorado's sample check10 of 10 (5 verbatim sentences, 5 table figures)10 of 10 (all verbatim sentences)10 of 10 (8 verbatim sentences, 2 table figures)10 of 10 (all verbatim sentences)10 of 10 (6 verbatim sentences, 4 table figures)10 of 10 (all verbatim sentences)
Note languageEnglish (the opening conclusion in Japanese)Japanese (some headings in English)JapaneseJapanese (finding headings and comments in English)Japanese (quotes with English translations)Japanese (quotes with English translations)

Dorado picked 10 findings from each note at even intervals, opened each source, and checked that the quoted sentence or the table figures are on the page. Headings, AI comments, and calculated rates were not checked.

10 of 10sampled findings in Astra's note with the quote or figures on the page
10 of 10sampled findings in Fable's note with the quote or figures on the page
10 of 10sampled findings in Opus's note with the quote or figures on the page
10 of 10sampled findings in Sonnet's note with the quote or figures on the page
10 of 10sampled findings in Sol's note with the quote or figures on the page
10 of 10sampled findings in Fable + Luna's note with the quote or figures on the page
0sampled findings with neither the quote nor the figures on the page (of 60)

Astra's note starts with a record written when its first helper stopped (marked "INCOMPLETE HANDOFF") and records from the pilot step, and the second helper's continuation follows. The final conclusion is at the top. In the sample, Astra's E08 gives p.11 for Nittoc's figures, but they are on page 14 of the PDF. In Fable + Luna's sample, the METI PDF (R3-F10) dropped the connection on the first fetch, and the quote was confirmed on the second.

Measured · elapsed time

Which model finished first?

12:13Astra, start to finish. Handed over at 10:27.
16:23Fable, start to finish. Handed over at 10:31.
29:14Opus, start to finish. Handed over at 12:34.
27:43Sonnet, start to finish. Handed over its second message at 09:05 on 15 Nov.
48:18Sol, start to finish. Handed over at 09:28 on 15 Nov.
55:16Fable + Luna, start to finish. Handed over at 11:31 on 15 Nov.
4.5×Fable + Luna took 4.5 times as long as Astra
Astra finished first, and Fable + Luna took more than four times as long

Time from prompt to answer. Astra and Fable started at 10:15 JST, Opus at 12:05, Sonnet and Sol at 08:37 and 08:39 the next day, Fable + Luna at 10:35 the next day

Opus's time includes the 22 seconds it waited for Minoru's answer on whether to research again. Sonnet's includes the 1 minute 24 seconds it waited for Minoru to approve the extra round. Sonnet's first message came 20 minutes 17 seconds after the prompt, at 08:58. Sol's time includes waiting for its three child tasks to finish. Sol reported about 46 minutes. Fable + Luna's time also includes waiting for its three child tasks, which ran 36 minutes 18 seconds together. Fable + Luna reported about 54 minutes.

Measured · tokens

How many tokens did each run use?

2.9MAstra's total in the logs. It reported 0.7M
5.3MFable's total in the logs. It reported 4.6M
11.7MOpus's total in the logs. It reported 9.8M
10.7MSonnet's total in the logs. It reported 9.9M
8.4MSol's total in the logs. It reported 2.9M
16.1MFable + Luna's total in the logs. It reported 15.4M
5.5×Fable + Luna's tokens against Astra's. Opus's were 4.0 times, Sonnet's 3.6 times, Sol's 2.9 times, Fable's 1.8 times

Fable planned 6 million tokens and reported 4.6 million at the end. The logs show 5.3 million. Astra set a cap of 12 million and reported 0.7 million, adding that its meter missed its helpers' calls. The logs show 2.9 million, and its two GPT-6.1 Sol helpers used 2.0 million of them. Opus planned 12 million and reported 9.8 million. The logs show 11.7 million. Sonnet planned 6 million and reported 6.8 million after three rounds. The extra round raised its plan to 8 million, and it reported 9.9 million at the end. The logs show 10.7 million. Sol set a cap of 12 million and reported 2.9 million, adding that it could not confirm its meter caught the child tasks. The logs show 8.4 million, and its three child tasks used 5.0 million of them. Fable + Luna planned 12 million and reported 15.4 million, 4.0 million of it for the main session. The logs show 16.1 million: 4.78 million for the main session and 11.36 million for its three child tasks.

Helpers used most of Astra's, Fable's, Sol's, and Fable + Luna's tokens. Opus's and Sonnet's main sessions used most of their own

Tokens by session, all six runs on the same scale

OpenAIAstra

2.94M

AnthropicFable

5.28M

AnthropicOpus

11.70M

AnthropicSonnet

10.69M

OpenAISol

8.41M

AnthropicFable + Luna

16.14M

Fable: main session (Fable 5.2) 2.08M, three Sonnet 5.5 helpers 0.91M, 1.07M, and 1.22M. Astra: main session (GPT-6 Astra) 0.95M, GPT-6.1 Sol helpers 0.49M and 1.49M. Astra's first helper stopped after 19 calls, so Astra gave the same task to a second one, which finished the three rounds. Opus: main session (Opus 5.5) 6.87M, three Sonnet 5.5 helpers 1.62M, 1.64M, and 1.57M. Opus's main session made 49 model requests. It prepared the page-fetching script, checked each round's quotes, and wrote the note itself. Sonnet: main session (Sonnet 5.5) 6.61M, four Sonnet 5.5 helpers 1.07M, 1.88M, 0.50M, and 0.63M. The fourth ran the extra round 5. Sonnet's main session made 41 model requests, and checked the quotes and wrote the note itself. Sol: main session (GPT-6.1 Sol High) 3.38M, child tasks on GPT-6 Luna High 2.79M and 1.36M, and on GPT-6.1 Sol Medium 0.87M. Each child task was a separate T3 thread that ran one round. Fable + Luna: main session (Fable 5.2) 4.78M, GPT-6 Luna High child tasks 1.29M, 4.54M, and 5.53M. The main session read only the child tasks' summaries and the saved quotes, not the pages, yet made 29 model requests at about 165K tokens each. The first Fable run's main session made 17 at about 122K each.

About nine in ten tokens in all six runs were earlier context read again (Astra 91.0%, Fable 89.5%, Opus 94.2%, Sonnet 92.0%, Sol 92.6%, Fable + Luna 94.9%). Astra wrote 17,583 output tokens, Fable 106,454, Opus 164,526, Sonnet 188,647, Sol 138,312, and Fable + Luna 214,630.

Estimated · cost

Which run was cheapest in dollars?

Models charge different prices per million tokens, so token counts alone do not compare cost. We divided each model's cost by its tokens in Minoru's usage dashboard (August to 15 November 2026) to get an effective rate, which already includes how much of that model's input is cached. Each session of each run is priced at its model's rate. These are estimates, not bills.

$1.84Astra's estimated cost, for 2.9M tokens
$2.91Fable's estimated cost, for 5.3M tokens
$3.74Opus's estimated cost, for 11.7M tokens
$2.35Sonnet's estimated cost, for 10.7M tokens
$0.83Sol's estimated cost, for 8.4M tokens
$5.26Fable + Luna's estimated cost, for 16.1M tokens
6.3×Fable + Luna's cost against Sol's
Sol cost least, and Astra, with the fewest tokens, came second. Fable + Luna cost most. The main session took more than 60% of every run's cost

Estimated cost per run, in dollars. Dark: the main session (planning, checking, writing the note). Light: the helpers (searching). Same scale for all

OpenAIAstra

$1.84

AnthropicFable

$2.91

AnthropicOpus

$3.74

AnthropicSonnet

$2.35

OpenAISol

$0.83

AnthropicFable + Luna

$5.26

Main sessionHelpers

The main session's share of cost: Fable + Luna 96% (30% of tokens), Astra 81% (32%), Fable 76% (39%), Opus 72% (59%), Sol 72% (40%), Sonnet 62% (62%). Only in Sonnet's run, where the main session and the helpers use the same model, do the two shares match.

Per million tokens, Astra's effective rate is 8.7 times Sol's

Cost divided by tokens in Minoru's usage dashboard (August to 15 November 2026, all use)

OpenAIAstra
$1.56
AnthropicFable
$1.06
AnthropicOpus
$0.39
AnthropicSonnet
$0.22
OpenAISol
$0.18
OpenAILuna
$0.02

Fable's and Opus's helpers ran at Sonnet's rate. Luna costs about 2 cents per million tokens, a tenth of Sol. In Sol's run, Luna's two child tasks read 4.15 million tokens for about 7 cents together. In Fable + Luna's run, Luna's three child tasks read 11.36 million tokens for about 20 cents together.

Source figures (cost / tokens): gpt-6-astra $2,992.89 / 1.92B; claude-fable-5-2 $343.20 / 324M; claude-opus-5-5 $4,413.96 / 11.3B; claude-sonnet-5-5 $404.67 / 1.84B; gpt-6.1-sol $2,826.86 / 15.8B; gpt-6-luna $0.51 / 28.9M

Hypothesis: a stronger model plans and improves, a cheaper one searches

This is Minoru's hypothesis, and the cost split of these six runs fits it. When the main session ran on a high-priced model, it took most of the cost even while reading the smaller share of tokens: Astra's main session read 32% of the tokens and cost 81%, Fable's read 39% and cost 76%. Helpers on a cheap model, on the other hand, read a lot for little money. Sol, the cheapest run, handed most of its searching to the low-priced Luna. Compare the two runs with Fable as the main session. Fable + Luna, which swapped the Sonnet helpers for Luna, cut the helpers' cost from $0.71 to $0.20. But its main session read 4.78 million tokens instead of 2.08 million, and its cost rose from $2.20 to $5.06. The total rose from $2.91 to $5.26, the time from 16 to 55 minutes, and the note's findings fell from 98 to 51. Each combination ran only once, though, Fable's two runs had different conditions in the request, and quality was not weighed against cost.

Measured · link check of the answers

Does each chat answer hold up as written?

After the runs, we opened every link cited in the answers and compared the page with the passage that cites it. A passage is backed when every figure, name, date, and condition in it is on the page. The sample check of the notes is in the research notes section.

9 of 10cited passages fully backed in Astra's answer
6 of 7cited passages fully backed in Fable's answer
5 of 6cited passages fully backed in Opus's answer
4 of 9cited passages fully backed in Sonnet's answer
7 of 7cited passages fully backed in Sol's answer
5 of 7cited passages fully backed in Fable + Luna's answer
0passages contradicted by their page, or dead links, in any answer
Sol is backed in all 7. Three answers have one passage only partly backed, Fable + Luna two, and Sonnet got the publication date wrong in 5 of 9

One square is one cited passage, checked on 14 Nov 2026, Sonnet, Sol, and Fable + Luna on 15 Nov

OpenAIAstra

9 of 10 backed

AnthropicFable

6 of 7 backed

AnthropicOpus

5 of 6 backed

AnthropicSonnet

4 of 9 backed

OpenAISol

7 of 7 backed

AnthropicFable + Luna

5 of 7 backed

BackedPartly backedNot backed

Fable put 7 links on 7 passages, and Opus 6 links on 6 passages. Astra put 11 links on 10 passages: one sentence cites both the JES earnings release and the MonotaRO results presentation. Sonnet put 9 links on 9 passages across its two messages. Sol and Fable + Luna each put 7 links on 7 passages.

What the check caught

  • OpenAIAstra: a rate that does not match its source.It puts data center services at about 8.2% a year. Fuji Keizai's release gives a 60.0% rise from 2024 to 2030, which works out to about 8.1%. The release also does not say the figures cannot be added to cloud. That is Astra's own conclusion.
  • AnthropicFable: a label its source does not use.It calls MIDAC Holdings an industrial waste company. The cited page backs the ¥11.8 billion in sales and the 39.9% operating margin, but describes the business as waste disposal and recycling, without the term industrial waste.
  • AnthropicFable: a comment that drops a later year.Its comment under the SEAJ figure says the same release forecasts a decline for FPD equipment. That holds for FY2026 and FY2027. The release forecasts FPD equipment up 27% in FY2028.
  • AnthropicOpus: the same label as Fable, not in its source.It calls MIDAC Holdings an industrial waste company. The cited earnings release backs net sales of ¥11,844 million, operating profit of ¥4,723 million, and the 39.9% operating margin, but does not contain the word waste. Opus cited the earnings release rather than an aggregator site, and the label problem stayed the same.
  • AnthropicOpus: "without reading the previous note" overstates it.The answer opens with this, but before researching, Opus read the opening of Fable's note (its summary of the answer), its coverage table, and its list of gaps. It told its helpers not to open the note.
  • AnthropicSonnet: wrong publication dates.All 9 quotations, figures, and calculated rates are on their pages. But 5 source lines get the publication date wrong. Three (GII, Tokyo Shoko Research, and Yano's single-serving food release) say the page gives no date, but the pages are dated. Two (complementary finance and renewable-energy maintenance) give the year in the survey's title (2025, 2024), but the releases came out in January 2026 and February 2025.
  • AnthropicFable + Luna: a publication date and a section that do not match the page.All 7 quotations, figures, and calculated rates are on their pages. But 2 source lines are wrong. The Impress Research Institute line says the page gives no publication date, but the page shows the report's issue date, 2026/03/27. The Yano retail line places the quote in "1. Market overview", but it sits in "3. Outlook". Unlike the page Fable and Opus cited, the MIDAC earnings release it cited uses the term industrial waste, in its description of the collection and transport business.
  • OpenAIAnthropicAnthropicAnthropicOpenAIAnthropicAll six: token reports that were too low.Fable reported 4.6 million tokens, Astra 0.7 million, Opus 9.8 million, Sonnet 9.9 million, Sol 2.9 million, and Fable + Luna 15.4 million. The logs show 5.3 million, 2.9 million, 11.7 million, 10.7 million, 8.4 million, and 16.1 million. Sol added that it could not confirm its meter caught the child tasks.
Compared · what each answer found

Where the six answers agree, and where they part.

TopicFableAstraOpusSonnetSolFable + Luna
Battery storage stations (the same Yano release)45.3% a year from FY2024 to FY2030, calculated. Notes that the price cap on its revenue source is expected to fall by more than halfAbout 45.3% a year from FY2024 to FY2030, calculated. Notes the same price cap41.4% a year from FY2025 to FY2030, calculated. Does not mention the price capNot in the answer. Mentions only the maintenance of grid-scale storage batteries, within renewable-energy maintenanceNot in the answer. The note leaves grid-scale batteries out of the comparison as a conditional capacity outlook41.4% a year from FY2025 to FY2030, calculated, via a news report. States that it is the market for revenue from operating the batteries, not the value of installations
Funeral services (Yano, 2025→2034)0.2% a year, calculatedAbout 0.17% a year, calculatedNot in the answerNot in the answer. States that no forecast for end-of-life services was foundNot in the answer. The note cites Tear's higher sales and 30.2% lower operating profit as counterevidenceNot in the answer. The note gives 0.2% a year from 2025 to 2034, calculated
AIDomestic AI spending, 36.0% a year (IDC)Installation, operation, and maintenance of AI infrastructure, 19.4% a year (IDC)Domestic AI systems market, 25.6% a year (IDC, 2024→2029)Domestic AI systems market, 25.6% a year (IDC, 2024→2029), alongside Fuji Chimera's 14.8% and JEITA's 47.2% for generative AINamed as part of the big picture, with no AI figure in the answerGenerative AI, 41.9% a year (Fuji Keizai, FY2024→FY2028, calculated). The note adds the whole AI market at 17.2% a year, calculated
CloudCloud infrastructure, 18.0% a year (Yano)Public cloud, 17.6% a year (IDC)Named as a growth field, with no figurePublic cloud, 16.3% a year (IDC). Fuji Chimera's figure is 10.4% a year, calculatedIaaS and PaaS, 14.6% a year (ITR, FY2024→FY2029). Notes price increases and demand concentrating on the largest providersIaaS and PaaS, 14.9% a year (ITR, 2026→2029, calculated, via a news report)
Niches that make moneySingle-company operating margins: MIDAC 39.9%, Sho-Bond 23.4%, Nihon M&A Center 37.3% (aggregator site)JES (elevator maintenance) 19.1%, MonotaRO 14.7% (calculated from earnings documents)Calculated from earnings releases: OBIC 65.7%, MIDAC 39.9%, Nihon M&A Center 37.3%, Daiseki 20.3%, JES 19.1%. Compared with 4.0% for all industries (MOF corporate statistics)By industry only: business-use machinery 15.3%, real estate 13.6% and others (MOF corporate statistics, ordinary margin), plus listed companies' medians by sector (EDINET DB). States that no margins were found for single nichesFrom IR documents: JES 19.5% (latest first half), Daiei Kankyo 25.3%. States these are neither industry averages nor new entrants' marginsFrom earnings documents: MIDAC 39.9% (calculated), Daiei Kankyo 25.3%, Sho-Bond 23.4%, JES 19.1%. States these are not industry margins
Other nichesCall center AI, 31.7% a year, plus a table of growth rates for 11 more industriesHalal testing, about 21.7% a year; data center services, about 8.2% a yearSpot work brokerage up 22.5% in one year (Yano). Timee's operating margin 18.1% (six months)Complementary finance, 32.6% a year; logistics robotics, 20.5% a year, calculated; drone business, 13.9% a year; single-serving food, 7.2% a year, calculatedMedical information system peripherals, 21.3% a year; logistics robotics, 19.1% a year (FY2027→FY2030); container data centers, about 41.7% a year; penetration testing, 12.4% a year (all calculated)Drone business, 13.9% a year (stated by the source). The note adds container data centers 41.7%, next-generation logistics robots 33.1%, and dementia-related products about 31% a year (all calculated)
Evidence against growthHome-care provider bankruptcies at a record for the third year (Tokyo Shoko Research)An aging population is not enough: the funeral market is almost flatRobot orders up 25.7%, but domestic shipments down 10.8% (Japan Robot Association)Care at 2.8% a year, staffing at 3.34% a year, pets, and reuse, as evidence against high growthEven if the market grows, small providers will not necessarily profit the same way (ITR)Domestic retail down about 14% from 2022 to 2030 (−1.9% a year, calculated). The note adds falling new housing starts, and flat funeral and construction markets
The 2026 to 2029 windowLists forecasts with different periods and labels each periodStates that it found no CAGR aligned to 2026 to 2029States that IDC's rate covers 2024 to 2029, not 2026 to 2029Lists forecasts with different periods and says to discount the gap with IDC for differences in definition and periodStates that it found no CAGR matching 2026 to 2029 and shows nearby periods as they areStates that only the IaaS and PaaS forecast matches the period exactly, and labels the period of each of the others

All three found the same Yano release on battery storage stations, but Opus calculated from a different base year (FY2025). Fable lists 18 growth rates. Astra narrows to 5 growth markets and 2 companies with actual profits. Opus spends 2 of its 6 findings on company and all-industry margins, and takes the margins from earnings releases. Sonnet leaves the battery storage release out of its answer, gives profit by industry, and says it found no margins for single niches. Sol calculates the same logistics robotics release from FY2027 (19.1% a year), and Sonnet from FY2024 (20.5% a year). Fable + Luna found the same battery storage release through a news report and calculated from FY2025 (41.4% a year), as Opus did. All four of its profit companies appear in other runs too: MIDAC and Sho-Bond in the first Fable run, JES and Daiei Kankyo in Sol's.

Measured · tokens per unit of output

What did one unit of research cost?

Each panel has its own scale. The first two count the research notes, and Fable used the fewest tokens. The other four count the chat answers. Astra used the fewest in three of them, and Sol the fewest per minute. The gap is widest per fully backed passage.

Tokens per finding in the note

OpenAIAstra
267K
AnthropicFable
54K
AnthropicOpus
150K
AnthropicSonnet
223K
OpenAISol
525K
AnthropicFable + Luna
317K

Against Astra: Fable 0.2 times, Opus 0.6 times, Sonnet 0.8 times, Sol 2.0 times, Fable + Luna 1.2 times

Tokens per distinct link in the note

OpenAIAstra
210K
AnthropicFable
34K
AnthropicOpus
119K
AnthropicSonnet
98K
OpenAISol
525K
AnthropicFable + Luna
359K

Against Astra: Fable 0.2 times, Opus 0.6 times, Sonnet 0.5 times, Sol 2.5 times, Fable + Luna 1.7 times

Tokens per character of the answer

OpenAIAstra
1,283
AnthropicFable
1,832
AnthropicOpus
4,608
AnthropicSonnet
2,648
OpenAISol
2,902
AnthropicFable + Luna
4,808

Against Astra: Fable 1.4 times, Opus 3.6 times, Sonnet 2.1 times, Sol 2.3 times, Fable + Luna 3.7 times

Tokens per cited link

OpenAIAstra
367K
AnthropicFable
755K
AnthropicOpus
1,949K
AnthropicSonnet
1,188K
OpenAISol
1,201K
AnthropicFable + Luna
2,306K

Against Astra: Fable 2.1 times, Opus 5.3 times, Sonnet 3.2 times, Sol 3.3 times, Fable + Luna 6.3 times

Tokens per fully backed passage

OpenAIAstra
326K
AnthropicFable
881K
AnthropicOpus
2,339K
AnthropicSonnet
2,673K
OpenAISol
1,201K
AnthropicFable + Luna
3,229K

Against Astra: Fable 2.7 times, Opus 7.2 times, Sonnet 8.2 times, Sol 3.7 times, Fable + Luna 9.9 times

Tokens per minute

OpenAIAstra
240K
AnthropicFable
323K
AnthropicOpus
400K
AnthropicSonnet
386K
OpenAISol
174K
AnthropicFable + Luna
292K

Against Astra: Fable 1.3 times, Opus 1.7 times, Sonnet 1.6 times, Sol 0.7 times, Fable + Luna 1.2 times

What the extra tokens bought

Fable's research note holds 98 quotes from three rounds of searching, and a script matched each one to its saved page. Opus's note holds 78 findings, records where on each page a quote sits (its section path), and takes margins from 10 companies' earnings releases. Sonnet's note holds 48 findings (4 of them overflow records) and separates, in its table, the rates publishers stated from the rates it calculated. Astra's note holds 11 findings. The answers run 2,884 characters for Fable, 2,538 for Opus, and 2,289 for Astra. Sonnet's two messages total 4,038. Sol's note is short, with 16 findings, and separates growth forecasts from company profit records in its tables. Sol's answer runs 2,896 characters. Fable + Luna's note holds 51 findings. A script matched its 55 quotes to the saved pages, and the note marks 8 findings as partly backed. Fable + Luna's answer runs 3,358 characters.

All six answers

All six answers, with their numbers.

CheckOpenAIGPT-6 Astra MediumOpenAIAnthropicClaude Fable 5.2 MediumAnthropicAnthropicClaude Opus 5.5 HighAnthropicAnthropicClaude Sonnet 5.5 HighAnthropicOpenAIGPT-6.1 Sol HighOpenAIAnthropicClaude Fable 5.2 Medium + GPT-6 Luna HighAnthropic · OpenAI
Tokens used (logs)2,936,4585,284,70511,696,24910,692,1558,405,50616,144,190
Estimated cost (USD)$1.84$2.91$3.74$2.35$0.83$5.26
Tokens the model reported0.7M (helpers not counted)4.6M9.8M9.9M (6.8M after three rounds)2.9M (child tasks possibly not counted, as it noted)15.4M (main session 4.0M)
Token plan or cap12M cap6M plan12M plan6M plan, 8M with the extra round12M cap12M plan
Elapsed time12 min 13 s16 min 23 s29 min 14 s27 min 43 s48 min 18 s55 min 16 s
Helpers2 GPT-6.1 Sol threads (the first stopped after 19 calls)3 Sonnet 5.5 agents3 Sonnet 5.5 agents4 Sonnet 5.5 agents (the fourth for the extra round 5)3 T3 child tasks (2 GPT-6 Luna High, 1 GPT-6.1 Sol Medium)3 T3 child tasks (all GPT-6 Luna High)
Answer length2,289 characters2,884 characters2,538 characters4,038 characters (two messages)2,896 characters3,358 characters
Cited links (distinct)876977
Websites cited555556
Cited passages checked1076977
Fully backed965475
Partly backed111502
Not backed or dead000000
Research note11 findings (plus 3 pilot records)98 quotes78 findings48 findings (4 of them overflow records)16 findings51 findings

Answer length is the character count of the Japanese original, without link addresses. Sonnet's two messages were counted one at a time and added.

All six outputs

Actual results

The research note and the final answer from each run, as produced. Switch the model, and switch between the note and the answer, to compare them. Links open the cited pages in a new tab.

AI translation of the Japanese original by Opus 5.5. Figures, links, and structure follow the original, which is the record. Read the original in Japanese

Research notes are shown as written, untranslated: Fable's and Opus's in Japanese, Astra's in English, Sonnet's in Japanese and English, and Sol's and Fable + Luna's in Japanese with English translations of their quotes.

  • Looking at both market size and growth rate, domestic cloud and the installation, operation, and maintenance of AI infrastructure are strong candidates. This is my judgment based on the primary sources below.
  • Among niches, watch halal testing and elevator maintenance. The first rests on a market growth forecast and the second on a company's actual profits. I did not confirm both "high growth and high profit" for either.
  • Battery storage station operation grows fast, but profit forecasts need caution. The price rules of electricity trading determine revenue.
  • I could not find CAGRs aligned to 2026 to 2029. The figures below are forecasts for other periods that include those three years. Figures for different periods cannot be turned directly into a ranking for the next three years.

Five fields to look at as growth markets

FieldForecast CAGRForecast periodMarket size and judgment
Battery storage station operationAbout 45.3%, calculatedFY2024 to FY2030¥45.0 billion→¥424.0 billion. Operator revenue from electricity trading, not battery manufacturing. Watch the price rules. Yano Research Institute, 2026-01-20
Halal testingAbout 21.7%, calculated2025 to 2035Forecast to grow 7.1 times. Still a small market of ¥1.0 billion in 2035. Fuji Keizai, 2026-05-14
Installation, operation, and maintenance of AI infrastructure19.4%2025 to 2030¥95.7 billion→¥232.0 billion. Worth watching as demand for professional services. IDC, 2026-04-08
Public cloud17.6%2025 to 2030¥4.4930 trillion→¥10.0962 trillion. Combines size and growth rate. IDC, 2026-03-12
Data center servicesAbout 8.2%, calculated2024 to 2030¥6.8189 trillion in 2030. Its market definition overlaps with cloud, so the two cannot be added together. Fuji Keizai, 2026-05-11

Calculated values were computed from the published multiple or the start and end points, as CAGR = (end ÷ start)^(1 ÷ years) − 1.

AI-adjacent services are especially worth digging into. IDC explains that equipment growing more complex, including liquid cooling, increases demand for professional services. My reading is to look not only at investment in the equipment itself but for companies that provide operation and maintenance after installation. However, this research did not confirm those providers' profit margins. IDC "Expansion of the AI Infrastructure Services Market and Changes in Competitive Dynamics", 2026-04-08

Halal testing is a candidate to look at as an added service for food testing companies. The original reads "● Halal testing ¥1.0 billion (7.1 times)". It covers the "2035 market forecast (compared with 2025)". The AI's judgment is that this is less a market for building a large standalone business than one for companies that can use their existing testing equipment and customer base. Testing alone cannot confirm every condition of halal certification. Fuji Keizai, 2026-05-14

Battery storage station operation cannot be chosen on growth rate alone. Yano's January 2026 report states that "the price cap in the balancing market is expected to be changed to less than half of the current level from FY2026". This is the outlook as of that report, and I have not treated it as the current, confirmed rule. Market growth and the profitability of each facility need to be checked separately. Yano Research Institute "2. Featured topics", 2026-01-20

Examples that confirm "actually making money"

Field and companyResultsOperating marginHow to read it
Elevator maintenance and modernization: JESConsolidated, fiscal year ended March 2026. Net sales ¥57.6 billion, operating profit ¥11.0 billionAbout 19.1%Net sales up 16.7% from the previous year, operating profit up 27.7%. A company example that confirms growth and profit. Not an industry-wide CAGR.
Procurement of supplies for factories and offices: MonotaRONon-consolidated, FY2025. Net sales ¥322.8 billion, operating profit ¥47.4 billionAbout 14.7%An example of profit in a business that supports procurement of consumables and similar items. Non-consolidated results, not an industry average.

Sources: JES "Consolidated results for the fiscal year ended March 2026" earnings release, 2026-05-14, and MonotaRO's FY2025 results in "2026 business plan, non-consolidated plan" results presentation, 2026-05-07. Margins were calculated as operating profit ÷ net sales.

If it were up to me, I would make equipment services with ongoing contracts, close to elevator maintenance and modernization, the next niche to research. JES's results are grounds for researching in that direction. However, they do not show that other maintenance businesses have the same margin.

Cautions when looking at the big picture

Higher demand from an aging population is not enough to pick high-growth industries. The funeral market is forecast to go from ¥1.8633 trillion in 2025 to ¥1.8915 trillion in 2034, a calculated CAGR of about 0.17%. It does not support the judgment that "the funeral industry as a whole will grow a lot because more people are dying". Yano Research Institute "3. Outlook", accessed 2026-11-14

From these sources, it is useful to look separately at growth in AI and cloud spending, businesses that balance electricity supply and demand, maintenance that keeps equipment in use longer, and specialized testing for specific food uses. On the other hand, how much attention a market gets ("nobody is watching") and high profits across whole niche industries remain unconfirmed.

The research ended after 3 rounds.

  • Round 1: checked 12 fields, separating domestic markets from overseas sales and CAGR from year-on-year change.
  • Round 2: strengthened the evidence on cloud, AI-adjacent services, and food testing.
  • Round 3: checked companies' actual profits and the limits set by prices and costs.

Unresolved: CAGRs for a common 2026 to 2029 period, checks of the original sources for AI software and cybersecurity, and an independent macro forecast. Source excerpts and the audit record are saved in the research note.

The meter showed 0.7M tokens of a 12M budget, 10 steps, and about 12 minutes elapsed. However, it did not capture every call, including those by delegates, so total usage cannot be determined.

The research note link at the end of each answer opens that note's download. One malformed link in Opus's note, which has no target, is shown as plain text. Sonnet's answer is two chat messages, and Dorado added the heading between them. The link to the quote audit in Sol's answer is shown as plain text, because that file is not published. The Fable + Luna tab is the second run with Fable 5.2 as the main session.

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.