Tokens: Neon 4.6M, Fable 7.64M, Astra 2.73M

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

GoogleNeon
4.6M, not reported
AnthropicFable
7.64M, reported 6.9M
OpenAIAstra
2.73M, reported about 2.4M
← Library · Model comparisons Model comparison · web search

Gemini 4 Neon vs Claude Fable 5.2 vs GPT-6.2 AstraResearching ways of communicating that come across as kind, polite, and competent in Japanese business

We tested how far Google's Gemini 4 Neon goes on the same research request. We compared it with Anthropic's Claude Fable 5.2 and OpenAI's GPT-6.2 Astra. Between 09:46 and 09:49 JST on 20 November 2026, we sent the same request, written in Japanese, to three threads. Every model ran at the "high" reasoning setting. All three used the same research skill (d-search) with the instruction to keep costs down (go cheaper).

For important research where thorough coverage matters, we still recommend using more than one model.

20 Nov 2026

Our takeaways
  • For important research where thorough coverage matters, we still recommend using more than one model.
  • Neon's output text is the easiest to read and hits the key points.
  • Another option is to have multiple models run the research and then have Neon edit the results into a report.
Brief · where Neon sits

Where Neon sits compared with the other two.

Neon compared with the other two

Each row is one measure. Zero is at the left end of the line, and each point's position shows the value for that measure. The blue point is Neon, and the grey points are Fable and Astra. Scales differ by row, so compare point positions only within the same row.

NeonFable, Astra
GoogleNeon13 min 35 s, 4.6M tokens. No usage reported. Note uncommittedAnthropicFable17 min 45 s, 7.64M tokens. Asked once when it reached its budget. Also listed what it did not findOpenAIAstra6 min 40 s, 2.73M tokens. Asked no questions. Found 5 cases with the receiver's own words

Each model ran once.

Prompt

The prompt sent to all three runs.

As sent, in Japanese

日本のビジネス環境で、日本人が親切、丁寧、仕事ができる、理解していると感じるコミュニケーションスタイルにはどういうものがあるのかを、実際の事例をもとに調査して

$d-search go cheaper

English translation

Research, based on real examples, what communication styles make Japanese people feel that someone is kind, polite, competent, and understanding in a Japanese business environment

$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?

13 min 35 sNeonanswered at 10:04
17 min 45 sFableanswered at 10:07
6 min 40 sAstraanswered at 9:56
Work time

Each bar is the work time from sending the request to the last answer. Time the model spent waiting for our reply is not in the bar and shows as a dotted line.

GoogleNeon
waiting 48 s

13 min 35 s. Answered at 10:04. Asked whether it could continue saving, and got our reply 48 s later

AnthropicFable
waiting 13 s

17 min 45 s. Answered at 10:07. Reached its 4M-token budget and asked how to proceed, and got our reply 13 s later

OpenAIAstra

6 min 40 s. Answered at 9:56. Asked no questions

0 min5 min10 min15 min20 min25 min
Time spent waiting for our reply

Astra asked nothing and finished in 6 min 40 s. Neon took 13 min 35 s, and Fable took 17 min 45 s.

Neon asked once whether it could continue saving the research note, and we chose "continue and allow saving". Fable asked once when it reached its budget.

Measured · tokens

How many tokens did each run use?

4.6MNeonnot reported
7.64MFablereported 6.9M
2.73MAstrareported about 2.4M
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. The black tick is the figure the model itself reported last. The dashed line is the 4M-token budget that all three set before researching.

GoogleNeon
Main 4.02M

4.6M (main 4.02M, 0.58M in 2 child tasks on Flash). Not reported

AnthropicFable
Main 3.45M
4.19M

7.64M (main 3.45M, 4.19M in 4 child tasks on Sonnet 5.5). Reported 6.9M

OpenAIAstra
Main 2.33M

2.73M (main 2.33M, 0.4M in 1 child task on GPT-6.2 Astra). Reported about 2.4M

02M4M6M8M
Main sessionChild tasks (Neon: Flash. Fable: Sonnet 5.5. Astra: GPT-6.2 Astra)Budget (4M tokens for all three)The model's own report

All three set a budget of 4M tokens before they started researching. Astra used 2.73M, staying within 68% of its budget. Neon used 4.6M, 15% over budget, and Fable used 7.64M, 91% over.

Neon's answer gives no usage figure. At the end of its answer, it writes that it could not get a measured value because the meter script (meter.py) does not read Antigravity logs. Counting from the logs gives 4.02M for the main session and 0.58M across 2 child tasks on Flash, for a total of 4.6M.

Fable asked once how to proceed when it hit 4M (4.3M in the logs). Research was done and only writing remained. We chose "research the items not found too, then finish". Fable's final reported figure was 6.9M, 90% of the logged total. Astra reported about 2.4M, 88% of its logged total.

The share taken by child tasks also differed. Child tasks accounted for 55% of Fable's usage, 13% of Neon's, and 15% of Astra's.

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 19 cited passages in the three answers, 13 were fully backed. The remaining 6 were partly backed. None were unbacked, dead links, or unverifiable.

4 of 7Neonfully backed
4 of 7Fablefully backed
5 of 5Astrafully 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.

GoogleNeon
4
3

4 of 7

AnthropicFable
4
3

4 of 7

OpenAIAstra
5

5 of 5

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

In all 6 partly backed passages, the quoted text itself matched the page. What did not match was the heading or the note the model added under the quote.

  • GoogleNeon: in its criteria for a good email, it wrote that "concise key points" was "first place". The page gives only the figure 85.05% and does not state a rank. The 2025 survey trend added in its note is not on this page either.
  • GoogleNeon: in its note on unpleasant emails, it cited 45.85% from the 2024 survey. That figure is not on the cited page.
  • GoogleNeon: it presented the quote from the Sales Practices Survey as being about corporate customers as a whole. The quoted sentence is about customers who expect "deep relationship-building sales", and the "60% continuing 5 or more times" in its note is also for that group. The contrast "rather than explaining their own products" is not on the page either.
  • AnthropicFable: it described the evaluation of Sony Bank as the words of "judges". The page says both general consumers and judges evaluate, and does not state which group wrote this passage.
  • AnthropicFable: it gave the source location for the RENGO survey as "p. 1-2". The quoted passage is on page 4 of the PDF.
  • AnthropicFable: in its note on overly polite wording, it cited the Agency for Cultural Affairs' 2018 report. That report is not on the cited page.

Neon's 2024 figure and 2025 trend, and Fable's Agency for Cultural Affairs report, each have separate source links in their research notes. In the answers, the models wrote them under a quote from another page without giving those links, so they were graded as partly backed. This check covers only the links in the answers.

All 5 of Astra's cited passages are quotes from one individual and carry no percentage figures. Neon's and Fable's cited passages include survey figures and add more figures in their notes. The more points there are to check, the easier it is for a mismatch to appear. This is our reading.

Dates also had small discrepancies. Astra wrote dates not found on the linked page in 2 items: "April 2023" for Builnet and "1 December 2024" for the mock-email study. We did not lower the grade for dates alone.

There were 4 factual sentences without a source link in Neon, 4 in Fable, and 2 in Astra. This count is not part of the grading.

Compared · what each answer found

Where the evidence chosen by the three answers overlaps, and where it parts.

Types of evidence cited

The 19 passages cited in the answers, grouped by type of evidence. Each dot is one cited passage. For Neon and Fable, 5 of 7 were questionnaire surveys with percentage figures. For Astra, 4 of 5 were company cases with the receiver's own words.

Type of evidenceGoogleNeonAnthropicFableOpenAIAstra
Questionnaire survey (percentage figures)550
Research paper201
Third-party rating or user review020
Company case (receiver's own words)004
Total775
Evidence in the answers

The items of evidence that appeared in the three answers. A filled dot means the model cited that item as a source. A ring means it only mentioned the item in a note or summary.

Evidence written in the answerGoogleNeonAnthropicFableOpenAIAstraTotal
Email surveys (Japan Business Mail Association)
Top unpleasant content: "not answering the question" 44.29%, 2nd "overbearing phrasing" 37.3% (Business Mail Fact-Finding Survey 2026)2
"Overly polite to the point of seeming insincere" (45.52%) is a clue that AI wrote the email (same survey)2
79.48% do not tell the sender even when they feel unpleasant (same survey)2
A good email has "key points summarized concisely" 85.05% (same survey)1
Top unpleasant email issue: "lacking necessary information" 45.85% (same association, 2024 survey)1
Over 60% feel a reply is slow if it does not arrive within 24 hours (same survey)1
Other surveys
Proactive horenso before being asked and fast responses build a manager's trust (Persol Research and Consulting, Kyushu University, 2025)1
What corporate clients expect of sales: "presenting hypotheses for value creation" and "organizing diverse internal opinions" (Persol Research and Consulting, Sales Practices Survey 2025)1
1st place for "a competent person" is "fast work and responses" (Biz Hits, 2021)1
Sales reps trusted by executives: considers issues together 51.4%, fast to respond 49.5% (Softbrain Service, 2023)1
Unforgivable complaint handling: passed around 41.8%, spoken to as if at fault 35.3% (RENGO, 2017)1
1st place for a dependable manager is "listens to me" 32.6% (Biz Hits, 2025)1
Research papers and government reports
Propose candidate dates yourself in a rescheduling email (Yokokawa, 2023, Journal of Japanese Language Teaching, No. 185)1
Point out a business partner's mistake without flatly asserting it, using "my understanding is" and "I think" (Yamashita and Ono, University of Tsukuba)1
In a mock email, "felt politeness" from a thank-you for arranging extra items (Ikeuchi, 2024)1
Honorifics can "put up a wall between people" (Agency for Cultural Affairs, 2018 report)1
Third-party ratings and user voices
Evaluation of Sony Bank's desk: "accurately grasping the customer's intent" and "proactive provision of information" (HDI-Japan, 2026)1
Auto insurance user review: "handled it kindly" (Oricon, 2015)1
Company cases (receiver's own words)
Builnet: a colleague sent the documents, and she "felt the warmth of colleagues helping one another"1
Osaka Energy Service: on the proposal to keep a page slated for removal, "a professional's perspective really is different"1
Tsubakimoto Kogyo: the salesperson's phrase "In short, your company is a problem-solving company" clicked right into place1
Frontline Works: weekly meetings made clear progress, what was missing, and the next tasks1
Cited as a sourceMentioned in a note or summary

0 items of evidence used by all three, 3 used by two (all Neon and Fable), and 19 used by only one (Fable 8, Neon 6, Astra 5).

Neon and Fable started from the same survey: the "Business Mail Fact-Finding Survey 2026" by the Japan Business Mail Association. The two answers quote the same two passages from it. One is that the top reason receivers felt unpleasant was "not answering the question" (44.29%). The other is that being "overly polite to the point of seeming insincere" (45.52%) serves as a clue that an email was written by AI. Their conclusions were close too: both pointed not to the volume of honorifics, but to answering what was asked and responding quickly.

After the email survey, the three diverged.

  • GoogleNeon: cited two research papers: proposing candidate dates yourself in a rescheduling email (Yokokawa, 2023, *Journal of Japanese Language Teaching*, No. 185) and pointing out a business partner's mistake without flatly asserting it, using "my understanding is" and "I think" (Yamashita and Ono, University of Tsukuba). It also used two surveys from Persol Research and Consulting.
  • AnthropicFable: alongside surveys, it used a third-party evaluation of a customer support desk (HDI-Japan's rating of Sony Bank), a user review (Oricon), and a consumer survey on complaint handling (RENGO). It added a "What was not found" section and stated that it did not find company cases showing a staff member's wording together with how it was rated.
  • OpenAIAstra: used no percentage figures at all. It chose five cases where receivers described their evaluation in their own words: four named company cases (Builnet, Osaka Energy Service, Tsubakimoto Kogyo, Frontline Works) and one mock-email study. It placed an English translation under each Japanese quote.

None of the evidence in Astra's answer appears in either of the other two answers. Astra included the 85.05% "concise key points" survey result in its research note, but did not use it in its answer.

Astra did find material close to the kind of case Fable said it could not find: client interviews on web-production agencies' sites and an employee interview on a recruiting site. Even so, among the five cases, only one (Tsubakimoto Kogyo) includes the exact words the staff member actually spoke.

What follows is our reading. For someone writing business emails or running meetings in Japan, how usable each answer is split as follows.

  • OpenAIAstra: gives phrasing examples for each situation. For example, instead of ending at "things are going smoothly", saying "The body text is finished. What remains is checking the price table. If you can confirm the amounts by Thursday, we can publish on Friday." Astra wrote these examples itself; they are not quotes. Astra notes that the examples do not represent the preferences of Japanese people as a whole.
  • GoogleNeon: pulled wording actually used from research papers, such as adding "if another time outside the options above works better for you" after listing candidate times, and pointing out a deadline mistake with "My understanding is that the agreed deadline was today." Because it also includes survey percentages, it works as material for explaining the point inside a company.
  • AnthropicFable: put the behaviors for the four impressions in a table and stated the target group and limits for each survey. It gives no phrasing examples. It fits readers who want to verify what each survey does and does not measure.
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 item of evidence, the note holds the source's original text and where in the source it appears. You can read all three notes in full under "Actual results" below.

The number of evidence items in each run's research note, and the note's length in characters.

  • GoogleNeon: 16 evidence items, 214 lines, 18,487 characters, 12 source links (no duplicates)
  • AnthropicFable: 33 evidence items, 398 lines, 22,768 characters, 20 source links
  • OpenAIAstra: 10 evidence items, 144 lines, 8,167 characters, 10 source links

Neon ran its own script to check its 16 quotes against the original text, and reported finding and fixing one single-character difference. It did not commit the note. A file showing that Git was in the middle of another operation (.git/index.lock) had been left behind, and Neon noted this on the first line of its answer.

Fable reported adding one paragraph with this run's measurements and a caution to the documentation of the skill that sets research budgets. That edit was not part of the request.

All three answers

All three answers, with their numbers.

ItemGoogleGemini 4 Neon HighGoogleAnthropicClaude Fable 5.2 HighAnthropicOpenAIGPT-6.2 Astra HighOpenAI
Tokens used (logs)4,602,3417,644,3572,732,888
Tokens the model reportednot reported6.9Mabout 2.4M
Token budget4M4M4M
Work time13 min 35 s17 min 45 s6 min 40 s
Questions along the wayOnce (checking whether to save, waited 48 s)Once (budget check, waited 13 s)None
Child tasks2 on Flash4 on Sonnet 5.51 on GPT-6.2 Astra high
Answer length (Japanese characters)5,1283,9023,155
Evidence cited in the answer775
Cited passages checked775
Fully backed445
Partly backed330
Not backed or dead link000
Factual sentences without a source link442
Evidence items in the research note163310
Research note commitNo (.git/index.lock remained)No (answer states uncommitted)Not stated in the answer
All three 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.

Done: Finished researching communication styles that come across as "kind, polite, competent, and understanding" in Japanese business settings and saved the research note (only git commit and push were not run because .git/index.lock dated 2026-11-20 10:01:23 remained).

In conclusion, the style evaluated as "kind, polite, competent, and understanding" in Japanese business settings is determined not by the volume of honorifics, but by a combination of structure that reduces the receiver's burden of reading, thinking, and replying (concise key points, direct answers to questions, offering options and a way out, fast initial responses and proactive horenso [reporting, contacting, and consulting]) and phrasing that protects the other party's standing and face (pointing out mistakes without flatly asserting them, designing requests around the other party's convenience). Conversely, excessive politeness that feels out of place for the sender's role, or messages that fail to answer questions or lack needed information, quietly lower evaluations without being pointed out in roughly 70 to 80% of cases.

Main surveys and cases shaping the conclusion (7 items)

  1. In the Japan Business Mail Association's 2026 survey (1,293 respondents), "concise key points" was first place at 85.05% as the criterion for a good email

    "For what made a business email feel well written, 85.05% named 'key points are summarized concisely', and this evaluation standard is shared regardless of whether AI is used." Business Mail Fact-Finding Survey 2026 | General Incorporated Association Japan Business Mail Association › Silent evaluations determine business trust · published 2026-06-01 (updated 2026-06-30) · source AI: This is the 2026 survey by the General Incorporated Association Japan Business Mail Association. Consistent with the 2025 survey's tendency that concise key points and clear structure matter more than etiquette and politeness, it shows that the core of appearing competent is concise organization of key points. It does not show percentages for spoken conversation.

  2. In the Japan Business Mail Association's 2026 survey, the top unpleasant email issue was "not answering the question" (44.29%), and second was "overbearing phrasing or a condescending tone" (37.3%)

    "The top item that made receivers feel unpleasant was 'not answering the question' (44.29%), and second was 'overbearing phrasing or a condescending tone' (37.3%). What stands out is that matters of communicative attitude and sincerity rank higher than technical mistakes such as typos or forgotten attachments." Business Mail Fact-Finding Survey 2026 | General Incorporated Association Japan Business Mail Association › Failures lower your evaluation out of sight · published 2026-06-01 (updated 2026-06-30) · source AI: This shows that what most makes Japanese receivers feel someone is "not understanding or insincere" is not typos, but failing to answer questions (44.29%) and an overbearing tone (37.3%) (in the 2024 survey, the top unpleasant item was "lacking necessary information" at 45.85%). It does not directly measure sufficient conditions for an active "kind" rating.

  3. In Yokokawa's (2023) study of rescheduling emails (16 native speakers and 18 non-native speakers working in Japan), proposing candidate times and "offering compensation" were practiced as consideration to reduce the other party's burden

    "In the open-ended questionnaire, businesspeople who proposed candidate dates and times gave reasons such as: suggesting dates themselves cuts down the number of exchanges, avoids troubling the other party as much as possible, gives the other party the choice, and makes it easier for the other party to reply." Consideration Language Behavior of Native and Non-Native Japanese Speakers in the Structure and Content of Japanese Business Emails (Journal of Japanese Language Teaching, No. 185) › 4-2-1 Content mentioned from codes within structural elements · published 2023-08 · source AI: This shows real examples of offering 2 to 3 candidate times yourself to cut back-and-forth effort, while adding an offer of compensation such as "If another time outside the options above works better for you..." to avoid being pushy. The same paper analyzes how "I am terribly sorry to trouble you, but could you please adjust your schedule to one of the dates and times below?" can prioritize one's own convenience in content even while using honorifics. Impression experiments on the reader's side are left as future work.

  4. In Yamashita and Ono's (University of Tsukuba) utterance study (20 businesspeople in their 20s and 30s), native Japanese speakers avoided flat assertions by using "my understanding is" (ninshiki da) and "I think" (to omou) when pointing out a business partner's misunderstanding

    "(6) and (7) use expressions such as 'my understanding is' and 'I think' to carry the implication that 'I only think so, and it may not actually be the case', avoiding a flat declaration of error and thereby trying to soften the threat to the other party's negative face." On Considerate Expressions in Business Situations by Native and Non-Native Japanese Speakers › 4.2 Scene 2: Pointing out a mistake · no publication date stated (accessed 2026-11-20) · source AI: This shows that when correcting a business partner's mistaken deadline, native Japanese speakers did not use the textbook "-yoda" ("it seems") (0 people), and instead made polite corrections that protected the other party's face using "my understanding is" (3 people) and "I think" (2 people), such as "My understanding is that the agreed deadline was today." Sample size is 10 per group, with written responses.

  5. In the Persol Research and Consulting and Kyushu University (2025) longitudinal study (5 organizations, 304 leaders and 1,848 members), proactive horenso before being asked and fast responses build trust, while going beyond instructions can backfire

    "To earn a manager's trust, proactive behavior such as reporting, contacting, and consulting at the right time before being asked, and prompt responses, were confirmed to be effective. On the other hand, even bold efforts intended to contribute to a leader can damage trust if they go beyond the scope of instructions, so care is needed in day-to-day communication." "Research on Trust Relationships between Managers and Subordinates" Announced - Persol Research and Consulting › For Members › Proactive behavior builds trust · published 2025-02-13 · source AI: This shows that timely horenso before being asked and fast responses link directly to being rated "competent and trustworthy" at work, along with the boundary condition that jumping ahead beyond agreed scope without alignment can hurt trust instead. The data comes from 5 organizations (with a high share of male leaders).

  6. In Persol Research and Consulting's "Sales Practices Survey 2025" (3,000 respondents total, including 1,000 corporate buyers), B2B customers valued "presenting hypotheses for value creation" and "organizing diverse internal opinions" rather than explaining one's own company

    "Among the behaviors expected of sales from problem formulation through the proposal stage, 'presenting hypotheses for value creation' and 'organizing diverse internal opinions' are emphasized in order to 'propose solutions from a sharp professional perspective'." "Sales Practices Survey 2025" Announced - Persol Research and Consulting › Survey Results Summary › 1. Expectations for relationship-building behaviors that engage deeply with customers · published 2025-11-11 · source AI: This shows that for B2B buyers making purchases of 1 million yen or more, the "understanding and professional" rating that leads to long-term business (60% continuing 5 or more times) comes not from introducing one's own products, but from presenting hypotheses about the customer's problems and organizing opinions inside the customer's company. Low-price and standard-product transactions are excluded.

  7. Counter-evidence: In the Japan Business Mail Association's 2026 survey, "phrasing that does not fit the person's role" (48.28%) and "overly polite to the point of seeming insincere" (45.52%) act as clues of AI authorship and unnaturalness, and 79.48% of unpleasant impressions lower evaluations without being pointed out

    "44.86% of people say they have noticed 'an AI must have written this' in an email received from someone else. The clues that give it away are human mismatches such as 'phrasing that does not fit the person's character or standing' (48.28%) and 'overly polite to the point of seeming insincere' (45.52%)." Business Mail Fact-Finding Survey 2026 | General Incorporated Association Japan Business Mail Association › Silent evaluations determine business trust · published 2026-06-01 (updated 2026-06-30) · source AI: This shows that overly humble honorifics and boilerplate out of step with the sender's role are seen through as unnatural rather than "polite". Because 79.48% in the same survey silently lower their evaluation without telling the sender even when they feel unpleasant, excessive honorifics and insufficient checking cause a loss of trust without the sender realizing it.

Summary by the 4 impressions (AI only)

  • "Competent" (sources: 1 and 5 above, plus Findings 2, 7, and 12 in the note): Concise key points and clear structure; a fast initial reply stating "by when I will answer" even when you cannot answer immediately; timely horenso before being asked (while avoiding solo actions beyond the scope of instructions).
  • "Kind" (sources: 3 above, plus Findings 7, 8, and 10 in the note): Proposing multiple candidate times yourself to save the other party back-and-forth effort while adding a way out (compensation) if none fit; cutting long apologies and unnecessary detail in already-apologized matters to spare the reader fatigue; noticing follow-up help that people around you could use right now.
  • "Polite" (sources: 3, 4, and 7 above, plus Findings 9, 11, and 15 in the note): Not only honorific forms, but content designed to leave the decision with the other party; phrasing that avoids flat assertions when pointing out a mistake, such as "my understanding is" or "I think"; honorifics that fit the role and situation without going overboard.
  • "Understanding" (sources: 2 and 6 above, plus Findings 3, 4, 13, 14, and 16 in the note): Answering the other party's question head-on with all needed information; offering problem hypotheses and organizing internal opinions for corporate clients rather than pitching your own company; recognizing the asymmetry that managers focus on "what you do" (horenso and execution) while subordinates focus on "how you engage" (explanation, listening, support); checking intent with options such as "A or B?" before rework when instructions are unclear.

Detailed note with all 16 quotes, English translations, and limits

Log of edits made during writing checks

  • The exact-match verification script detected a one-character difference in Finding 5 ("にのぼり、" -> "にのぼります。") and fixed it to exact match (16/16).
  • As an M3 and date check, changed the heading of Finding 7 to the survey item name itself, and removed an unlisted publication year in Finding 15, setting it to Not stated.

Tokens: actual count not obtained because meter.py does not measure Antigravity logs / budget 4M / about 12 main turns (+ searches from 2 stopped lightweight flash subagents) / about 17 min elapsed (including waiting for write-permission confirmation).

The research note link at the end of each answer opens that note's download.

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.