Stepping my toe into the agentic pool with Duet

I felt like a kid waking up on Christmas morning when I opened my phone to a notification that I had been taken off the Claude Duet waitlist - woo hoo! I had signed up for the waitlist when I saw the announcement on October 21. One of the caveats is the preview is limited to Pro and Max accounts; I have the Max account. Duet is Claude's new agentic feature, like ChatGPT's Dots. I only have a Plus subscription with ChatGPT so I don't qualify for Dots...yet. But I was fortunate to see Steve Little use Dots and it's pretty cool...like if Siri was smart.

First things first, I had to update the Claude app on my phone. I noticed after the update that a button appeared so you can have a conversation with Duet like Steve did with Dots. (BTW - Steve's Dot has a Silicon-Valley girl accent. I notice these things.) In trying out Duet's voice chat, the default voice it chose was a male British accent. I asked Duet where in Britain the accent was from and it would not give me a region, it referred to it as a "subtle British lilt". Yeah right Duet, go tell that to a British native! I am quite certain this "lilt" has a region and if I had to guess it's placement, somewhere around the middle, like Birmingham. I have been to England enough times to hear their regional accents, but if you try this feature, let me know your thoughts on the accent's region.  

I next asked Duet some more feature particulars and Duet replied, "Most AI tools wait for a prompt. Anthropic’s new app, Claude Duet, is built to do the opposite. It runs in the background, reports what needs attention, and works on tasks you assign." It also said that you cannot choose a model for Duet, it runs on Opus 5.5 and Anthropic gives you one token "Reset" you must use by Nov 30 if you tokenmax. ChatGPT's Dots uses Astra so I would blow through that in no time flat! I am thrifty by nature and that includes my tokens. 

Ok, kid in the candy store moment - what do I try first with Duet? Well, I'm a longtime genetic genealogist with an interest in epigenetics so I task Duet to compile a weekly summary of published scientific studies with keywords like "epigenetics", "methylation", "haplogroup", "mitochondrial" and "Y-chromosome". I give it two sites to query and Duet hits the wall just like the majority of informational sites have had to install from AI bots driving up their server costs. Not one to be easily thwarted, I find that both sites have smartphone apps and neither requires an account. I download both apps and give Duet permission to access those apps on my phone and do a test run. Worked like a charm! Doing a weekly summary is also token thrifty, I have only used 2% of my limits so far today. Please share your agentic genealogical uses as I would love to try more.

AI Glossary: "Benchmaxxing" - when developers game the system

What Is Benchmaxxing?

Benchmaxxing is AI-industry slang for optimizing a model to score well on benchmark tests rather than to be good at the capability those tests are supposed to measure. I first heard the term on The Neuron: AI Explained podcast, in the hosts’ coverage of Meta’s Muse launch. As the cybersecurity firm CrowdStrike put it, once a score becomes the goal, teams start optimizing for the benchmark instead of the underlying skill. A modern application of Goodhart’s Law: a measure ceases to be useful once it becomes a target.

Meta’s Benchmaxxing Reputation

In the frontier AI field, no company seems to be more associated with the term "benchmaxxing" than Meta. When Llama 4 launched, its benchmark charts suggested frontier-level performance, but developers found the real model disappointing, and “benchmaxxing” trended on Hacker News before launch week ended. The reputation followed Meta into its reorganized Meta Superintelligence Labs and the Muse model family. When Muse Spark debuted in April 2026 with charts claiming parity with the best models from Anthropic, OpenAI, and Google, developers testing it on launch day reported basic math errors, and one commenter claiming insider knowledge blamed Meta’s performance-review culture, which rewards showing good numbers.

Meta’s own safety disclosures added a twist. Apollo Research, a third-party evaluator, found Muse Spark showed the highest rate of “evaluation awareness” of any model it had tested: the model frequently recognized test scenarios as tests and reasoned it should behave honestly because it was being evaluated. Evaluation awareness is not benchmaxxing; one is the model noticing the test, the other is the developer optimizing for it. But both corrode the same thing: confidence that test behavior predicts real behavior. Meta concluded the finding was not a blocking concern for release.

What This Means for Genealogists 

AI companies do not measure benchmarks on genealogical skills. Something to keep in mind: a model that tops a math or coding leaderboard has not been tested on family history benchmarks like keeping two same-named men separate until evidence merges them, analyzing family tree errors, or admitting a record doesn’t exist rather than inventing a citation. That's why you need to proof for accuracy, In the end, don't let choosing an AI model based on benchmarks be the deciding factor as those could be gamed; choose a model that has the best tools and features to serve you.

  

Compiled by the Human, Claude Fable 5 and ChatGPT Sol 6.1

Super-easy Claude Design for Genealogy Slide Show Presentations

Claude Design - Design System slide deck

 Last August was my first go at using Claude Design to create my presentation for East Coast Genetic Genealogy Conference.* The layout and text were Claude's, and the background images were ChatGPT. Today, I decided to practice creating different templates and content with Claude Design. I had read that the tool could be a token drain, so I was mindful of that and used Opus 4.8 medium. Granted, I have a Max subscription so only hit 6% usage after three attempts on this session. 

Step one to create the slide deck, open claude.ai on your computer (not available in the phone app) and click "Design" on the left menu. I began with creating a "Design System" so click that box. Under "Reply" I prompted, "Create a slide deck with graphics for artificial intelligence and genealogy." I then used ChatGPT to compile a summary of one of the best beginning level AI YouTube presentations I have seen, "AI-powered Family History Research" by Colette Hokanson. Next, I had ChatGPT update the summary with current information, stats, etc, and create a doc which I downloaded. Going back to Claude, I chose the newly-created Design System from the drop down menu (it assigned it a name), uploaded the summary and asked it to create a slide deck and cite sources. Here is the final product. Probably took a total of 20 minutes to create in gathering the info, prompts, review and then generation.

*You can actually still sign up for the East Coast Genetic Genealogy Conference and stream sessions until December 31, 2026.

  

What it means to turn off the Training Data

I am a big fan of Steve Little's and Mark Thompson's podcast, "The Family History AI Show". Their show always makes me crave more so I began listening to other AI podcasts as well like "The Neuron: AI Explained"  A recent episode discussed Meta's launch of their AI "Muse". Grant Harvey and Corey Noles discuss the generous 100 million token allowance for free users. They read the terms of service and note that when you use the Meta Muse FREE account - to quote Grant, "They (Meta) can do anything they want with your data." Musing about Muse, I wondered, "What would they want to do with MY data?"

So of course, I posed that question to Claude. Already a Claude user, it has a history of my use for genealogy. Here is what Claude replied when I asked why it would want to train on my data:

"Your chats are valuable for that stage in a few specific ways. They are real rather than invented test prompts. They show niche, expert work (reading an 18th-century plat, citing a county court order book) that is thin on the open web. They capture multi-turn corrections, which is how the model learns what it got wrong. And they are increasingly scarce, since publishers and archives now block crawlers, so user conversations have become one of the few growing sources of fresh, high-quality text."

Regarding the "block crawlers" - I fall into the "publishers" category and while I share genealogical information freely, the crawlers drove up server costs so we had to install Cloudflare. At the time this was happening in 2021, we everyday folks had no idea that it was AI crawlers building their knowledge-base. 

Leaving the training data toggle switch on, allows Claude to use your data for five years. So I asked again, "Why me? And if I turned it off, would it delete my work?"

Muse billboard in rural Central California

Claude's reply: "The reason the “why me” feeling is reasonable is that the pool only works if it includes people like you. Millions of casual users asking for recipes and email drafts teach the model to do recipes and email drafts. The rare person transcribing colonial court records, reasoning through land grants, or correcting a bad citation teaches it something the other millions can’t. From a training standpoint your conversations are more useful than average precisely because they are unusual. That is a compliment to your work and also the best argument for keeping it out: the more distinctive the content, the less "lost in the crowd” it really is."

I had already toggled it off before I started this thread, but I just wanted to ensure this did not negatively impact my work because when items are deleted, they are permanently deleted after thirty days. If the setting is on, retention can be up to five years. Data already in a training run that has started, or in a model already trained, cannot be removed. (Anthropic Privacy Center)

A couple of caveats on Muse you should know: even with a free account, you still have to provide a payment card. From Mindwiredai: 

"You cannot start on the free tier without adding a payment card. Meta’s reasoning is that subscriptions kick in as usage rises, so the card has to be there first. You are not charged unless you upgrade or approve a purchase — but “free” and “give us a card” are two different sentences, and only one of them made the press release. Free tiers this large are never charity, and Meta hasn’t pretended otherwise. In a September 8 interview with journalist Alex Heath, Zuckerberg described the plan: Meta expects to eventually take “a very small cut” when Muse helps you make money, save money, or complete a purchase."

Some free account users have reported using Muse on the web and not having to provide a credit card. Do read the full terms of service so you are not caught unaware if the fees are implemented. From their TOS:

"Fees. Meta may decide to offer Muse for free for a limited period, may decide to allow a free trial from time to time, or may decide to charge for use of Muse. Meta reserves the right, in its sole discretion, to modify or discontinue offering Muse for free, including any free trials, at any time and without prior notice, unless where required by law."

Two things set Muse apart from Meta AI proper. From Engadget, "US users have no option to opt out of training and since December last year, the company has used Meta AI chats for ad targeting, also without a way to decline."

I had also turned off training data on ChatGPT last week. So where does that leave Muse and it's generous 100 million tokens? I guess it depends on how you feel about your content being used for training as to whether you decide to toggle it off. 

About Venice AI: Tokennomics for Privacy Nerds

Even before I ever subscribed to an AI app, I heard an ad for Venice AI on a podcast commercial over a year ago so I downloaded it to my phone. As I have been tokenmaxxing on my subscriptions as of late (not on purpose) I will be using Venice more. With the debut of Meta's Muse and it's large token allowance, I have thought more about privacy and just what am I giving to AI companies in exchange for using their products? With a free Muse account, Meta is compiling a dossier on it's users. Even "free" has a price. And even a paid account doesn't automatically mean you get to bypass the data brokers. Grok, for example, opts everyone in. OpenAI does too on free, Plus, and Pro accounts. I subscribe to ChatGPT Plus, their $20 a month plan. In researching "training data", I learned that even if you have a ChatGPT paid account like Plus or Pro, you are automatically opted in to having what you upload be used as "training data". So I opted myself out. Caveat -  even opting out does not opt you out if you provide feedback:

"Even if you have opted out of training, you can still choose to provide feedback to us about your interactions with our products (for instance, by selecting thumbs up or thumbs down on a model response). If you choose to provide feedback, the entire conversation associated with that feedback may be used to train our models."

So I decided to ask Venice AI what it has to offer for genealogy and what are the pros and cons of the model:  

Brickwall Research Plan by Venice AI on John Sisty

First, the actual numbers:

Venice publishes its tier limits directly, and they're stricter than most reviews claim. From Venice's own help documentation:

| Tier | Text prompts/day | Character prompts/day | Image requests/day |
|---|---|---|---|
| No account (anonymous) | 5 | 5 | 10 |
| Free account (email/wallet) | 10 | 5 | 16 |
| Pro | Unlimited | Unlimited | 1,000 |

Two footnotes matter enormously for genealogy: image prompt consumption varies by use case, and the Upscale & Enhance features can consume up to 20 requests per operation. Limits reset daily at 12:01 AM UTC.

Note the discrepancy: several third-party reviews cite "25 text + 15 image" for free. Venice's own page says 10 text / 16 images. Trust Venice's page. [Venice FAQ] 

What one unit of genealogy work actually costs

Genealogy is unusually expensive in prompts because it's verification-heavy — you rarely accept a first pass. Realistic costs on a free account:

| Task | Prompts |
|---|---|
| Transcribe one handwritten census page (vision) | 1, realistically 2–3 if the hand is bad |
| Translate one German/French/Spanish parish register entry | 1–2 |
| Extract structured fields (name, date, place, parents) from one record | 1 |
| Write an EE-style source citation | 1 |
| Draft a research plan for one brick wall | 1–2 |
| Cluster a DNA match list you paste in | 2–3 |
| OCR cleanup of a 5-page will (chunked) | 3–5 |
| Upscale one damaged ancestor photo | **up to 20 image requests** |
 
Three worked sessions

Session A — "One record, done properly." Free account, 10 text prompts.

1. Upload the 1880 census page → verbatim transcription
2. Second pass: flag every uncertain letter and why
3. Structured field extraction into a table
4. EE-format source citation
5. Cross-check the household against an index you paste in
6. Draft a one-paragraph research summary

That's 6. You have 4 left for corrections. Result: one record family taken to a defensible standard per day. Not a family line — one record.
 
Session B — "Batch transcription day." Free account, 10 prompts, 1 per page, zero correction budget. You'll realistically get 6–7 pages because you will need re-reads. A 30-page census district takes four to five days. This is the single biggest practical constraint on free-tier genealogy.

Session C — "Photo regeneration day." Free account, 16 image requests. Because a single Upscale & Enhance can burn up to 20 requests, one upscale can exhaust your entire daily image allowance and then some. Realistic output: two or three basic regenerations, or exactly one upscale if it runs long.

Session D — Anonymous. 5 text prompts. Enough for one transcription plus one citation. Nothing more.

Timezone hack worth knowing: the reset is 12:01 AM UTC. In US Eastern that lands at 8:01 PM EDT / 7:01 PM EST — so an evening session straddles the reset and you effectively double your daily allotment. Europeans (1:01 AM CET) get no such luck without staying up.

Give the Venice AI free account a test drive and comment if you have any successes. (Note: using the hyperlink will give us both credits for any Pro sign ups.)

For more on privacy in genealogy and AI, see the Coalition of Responsible AI in Genealogy craigen.org

 

ChatGPT for Genealogists: Models, Plans, Maps, and Privacy

(Graphic generated by Luna 6 Medium)

Two days ago, I hit my ChatGPT usage limit (aka "tokenmaxxing") while using Astra on Max. I had generated three maps and asked three follow-up questions when I hit the limit, so learned my lesson the hard way to toggle the model to save tokens. ChatGPT uses varying verbiage for usage; it does use the term "tokens"...sometimes...but weirdly, not most of the time.

The map I was generating was a map of where my ancestors lived in 18th century Westmoreland County, Virginia. I will eventually do a post about how to generate such a map, but ChatGPT still struggles to get the modern roads correct; so it is not quite there yet with accuracy. One of my cousins was generating a map using the free ChatGPT account and hit the usage limit which made me wonder about the differences between a free vs. paid account. My account is the $20 Plus plan. So I prompted ChatGPT to compile the following information about the different plans and suggested uses for genealogists. Lastly, I turned off the setting which allows what I enter to be used for "training data". Not that I am entering anything on living persons, etc. But just to be on the safe side. Note too how the setting is called, "Improve the model for everyone" - lol! Passive-aggressive guilt trip much, OpenAI?

Information checked October 14, 2026. Prices are in US dollars. Features, limits, and model access can change.

ChatGPT can help genealogists organize research, transcribe records, compare evidence, draft biographies, and create maps showing where ancestors lived. The model, the subscription plan, and the tools you can use are related, but they are not the same thing.

A model processes your request. Your plan affects which models and features are available and how much you can use them. Tools such as file uploads, data analysis, and image input help you work with source material.

Chat is suited to questions and discussion. ChatGPT Work is designed for larger tasks and reviewable results, such as a report made from uploaded records. Codex is intended for programming and other technical work. Most genealogists will begin with Chat or Work.

ChatGPT models made simple

Model names and availability can change. As of this review, GPT-6 is rolling out in Chat to eligible paid users. Free and Go users receive GPT-5.6 Luna as the default model. In GPT-6, the Instant, Medium, High, and Extra High thinking levels use GPT-6 Sol. The Pro thinking option uses GPT-6 Pro, powered by Astra. GPT-6.1 Sol is a separate option in Work and Codex, rather than ordinary Chat. Model access depends on your plan, account, workspace, and rollout status.

Model or option	What it is intended for	Possible genealogy task
GPT-5.6 Luna	Everyday questions and focused work	Organize notes or summarize a short transcription
GPT-6 Instant	Fast answers to everyday questions	Ask what a historical term in a record means
GPT-6 Medium or High	More involved analysis	Compare several records and identify unanswered questions
GPT-6 Extra High	The highest GPT-6 Sol thinking level	Review a complicated timeline for contradictions
GPT-6 Pro, powered by Astra	Difficult problems and longer tasks	Compare competing identity hypotheses using multiple records
GPT-6.1 Sol in Work or Codex	Complex work with attention to usage	Prepare a chronology or sourced biography from supplied documents

These are suggested applications, not a guarantee that one model will produce accurate genealogical conclusions. Check the current model menu in your account.

Free and paid plans

Here are the main plans and listed prices relevant to individual genealogists and small groups:

Plan	Listed price	What may matter to genealogists
Free	$0	Everyday text chat; file uploads, data analysis, image creation, and other tools have separate limits
Go	$8 per month	Higher limits than Free for eligible features such as file uploads, image generation, and data analysis
Plus	$20 per month	Access to additional models and Work; usage allowances still apply
Pro 100	$100 per month	Pro models, file uploads, deep research, and a larger included usage allowance
Pro 200	$200 per month	More included usage than Pro 100
Pro 500	$500 per month	Highest Pro allowance; includes eligible Ultrafast access
Business	$20 per user monthly with annual billing, or $25 monthly	For at least two users; includes shared workspace and administrative features

Prices and features are subject to change. Pro 100 and Pro 200 do not include Ultrafast. Pro 500 does.

Free users can continue everyday text chats, subject to safeguards, but tools such as uploads and data analysis have separate limits. Paid plans provide higher limits for eligible features. Pro offers larger allowances, but it is not unlimited access to every model or tool.

A genealogist who asks occasional questions may find Free sufficient. Go or Plus may suit someone who regularly works with uploaded documents or spreadsheets. Pro may be useful for frequent, demanding projects, but the higher price makes sense only if the additional capacity meets a real need.

Creating an ancestral migration map

Both Free and Pro users can ask ChatGPT to help create a map, though their access to file uploads, data analysis, and other tools differs. Think of the examples below as sample project sizes, not promised plan limits.

Free example	Pro example
Provide a short list of five verified locations for one ancestor	Upload a spreadsheet with locations for several generations
Ask for labeled points and arrows showing the order of documented residences	Ask for separate colors for family branches, labels, dates, and a legend
Review the map and make a small number of revisions as your available usage permits	Work through multiple revisions or request separate versions for printing and presentation
If you reach a tool limit, wait for the reset shown in your account	Check your usage before starting a large project; Pro still has an allowance

A clear prompt might say:

Create a migration map from the attached table. Label each ancestor, location, and year. Use verified coordinates and a geographic base map. Put uncertain locations in a separate list. Do not invent a route between documented residences.

A plotted map based on verified coordinates is different from an AI-generated picture that merely looks like a map. Check every place name and location against your sources. Treat generated historical images as illustrations, not evidence, and label them accordingly.

Tokens and usage limits

OpenAI uses the word tokens for small pieces of information that ChatGPT reads and writes. Your prompt, files, conversation history, tool results, and answer all contribute to token use. However, OpenAI does not publish one fixed token total for Free users or a single fixed token total for Pro users. So there is no reliable number to give for “Free tokens versus Pro tokens.”

OpenAI also describes subscription use with terms such as usage allowance, model allowance, and usage limits. In simple terms, an allowance is the amount of work included with your plan. You may have limits for a particular model, a time period, or a tool such as file upload or data analysis.

Term	Plain meaning	Genealogy example
Tokens	Pieces of information the model processes	A long probate file takes more processing than a short question
Usage allowance	The work included with your plan	Comparing a set of records and drafting a biography uses some of your included access
Model allowance	A limit for using a particular model	Astra may become unavailable after you have used its allowance
Tool limit	A separate limit for a feature	You might still be able to chat after reaching a file upload limit
Reset time	When access becomes available again	ChatGPT may show when you can use the affected model or tool again

The same number of messages can use different amounts of allowance. A short request to compare several deeds, tax lists, and probate records may require more work than a longer request to alphabetize names. OpenAI says usage depends on the model, context, reasoning settings, and tools used.

OpenAI publishes estimates for local Work and Codex messages on Plus, within a five-hour window. These are estimates, not guaranteed counts, and they do not apply as Free or Pro message limits.

Model	Estimated Plus messages in five hours
GPT-6 Astra	5 to 45
GPT-6.1 Sol	15 to 160
GPT-6 Sol	15 to 150
GPT-6 Luna	350 to 3,000

Weekly limits may also apply. Pro has no five-hour usage limit in Work and Codex, but it still has an included allowance. GPT-6 Pro in Chat has separate limits. Your actual available use depends on your plan, model, task, and settings. Check Settings → Usage for your allowance and reset times.

For example, asking Astra to read a group of deeds, resolve uncertain place names, compare conflicting evidence, and prepare a sourced explanation can consume more allowance than asking it to summarize one document. Your experience of reaching an Astra limit is consistent with OpenAI’s guidance: Pro users can reach model limits too. Depending on the feature and account, you may need to wait for a reset, switch to another available model, or use eligible credits. Switching models does not replenish a shared Work and Codex allowance.

To make usage go further, upload only the relevant records, state the exact question, and request the format you need. You might first ask for a transcription with uncertain readings marked, then ask for a comparison of the transcribed documents. Higher reasoning effort can use more allowance and does not always improve the result.

Privacy and uploaded family records

Yes, OpenAI may use content from personal Free and Pro accounts to improve its models when the relevant setting is on. This can include content from conversations and eligible uploads. The setting can be turned off on either plan.

To opt out, open Settings → Data Controls and turn off Improve the model for everyone. New conversations will not be used for model training after you opt out. Your chats can remain in your history. Turning this setting off does not delete saved conversations or files. If you submit feedback on a response, OpenAI says the conversation associated with that feedback may still be used to improve models.

CRAIGEN, the Coalition for Responsible AI in Genealogy, states its Privacy Principle this way: genealogists should take reasonable measures to safeguard private information when using AI. For family history work, consider removing names, exact birth dates, addresses, and other identifying details about living people when they are not needed. Take particular care with adoption records, health information, and DNA match lists. CRAIGEN’s privacy principle and practical privacy guidance offer further advice.

Whatever plan you use, check transcriptions against the original images, follow cited links to the underlying records, and verify names, dates, relationships, and map locations. CRAIGEN’s Accuracy Principle calls for checking AI-generated information against other records and acknowledging credible sources.

Sources

1. OpenAI, Use ChatGPT.

2. OpenAI, GPT-6 and other models in ChatGPT.

3. OpenAI, Model selection.

4. OpenAI, Pricing, tokens, and usage allowances.

5. OpenAI, Managing usage with GPT-6 Astra in Work and Codex.

6. OpenAI, About ChatGPT Pro tiers.

7. OpenAI, ChatGPT Free Tier FAQ.

8. OpenAI, Data controls in ChatGPT.

9. OpenAI, How your data is used to improve model performance.

10. OpenAI, Data analysis with ChatGPT.

11. Coalition for Responsible AI in Genealogy, Guiding Principles.

12. Coalition for Responsible AI in Genealogy, Key Action Points for Protecting Privacy.

Disclaimer: This post was compiled by Katherine Borges and ChatGPT. Katherine Borges has no paid affiliation with OpenAI. This post is for informational purposes only.

OpenAI DevDay 2026: What Dots, Spaces, and the New Tools Could Mean for Genealogists

OpenAI's DevDay keynote on September 29 included more than 20 announcements. Most were aimed at developers and businesses, but several of them line up well with the daily work of genealogists: long research projects, piles of records to transcribe, shared family trees, and sensitive DNA data. The Neuron podcast hosts reacted live, and I watched the replay with a genealogist's eye.

Dots: an assistant that keeps working

The headline announcement was Dots, an always-on personal agent built into ChatGPT. Dots has its own cloud computer and browser, connects to more than 4,000 apps through ChatGPT plugins, and keeps working in the background. OpenAI's demos showed Dots reviewing overnight messages, rebuilding a slide deck, rescheduling a conflict, and fixing software bugs. Dots is included in the plan and does not count against usage. It is available now to Pro, Business Premium, and Enterprise users.

Steve Little introduced me to Dots last week and showed me how to use it like an assistant, and that framing is the key. Think of a research assistant who drafts the correspondence, organizes the research log, summarizes what came in overnight, and reminds you what is waiting for your review. For a genealogist juggling several projects, that is a very different tool from a chatbot you have to prompt from scratch every time.

One personal observation: Dot's speech pattern sounds to me very "Silicon Valley Californian". In the keynote demo, Dots was upbeat and casual, even handing out wedding advice. Dots is like having an Astra-level smart Alexa or Google Nest that you can converse with; so you could talk through a brick wall problem on the drive home from the courthouse or archive and have Dots pick up the thread when you sit back down at your desk.

Spaces: a shared workspace for research teams

ChatGPT Spaces gives teammates and Dots a shared place to work on pages, files, and data. Someone can tag Dots in a comment and it gets to work. For genealogists, the obvious fits are family history societies, project volunteers, and cousins collaborating on a shared line. A space could hold a research log, a source list, and a draft write-up, with Dots keeping the log current while the humans make the judgment calls. OpenAI also said a collaborative presentation feature is coming, which could help with society talks.

Codex in the cloud: long jobs without babysitting

Codex now runs fully in the cloud, so a task started on a phone keeps running after the laptop closes. For genealogists who build their own scripts, this suits the long, tedious jobs: batch processing record images, cleaning and merging spreadsheets of census extractions, or analyzing large sets of DNA matches. You hand off the task and check on it later.

Speed and cost: Sol, Ultra fast, and cheap cached input

GPT-6.1 Sol is described as delivering close to Astra-level intelligence at about a fifth of the price, and cached input is discounted by 95 percent. Caching matters for genealogy because research is full of repeated context. When you work through a long run of court records or a family file, the same background material gets reread again and again, and cached input makes that far less expensive. Ultra fast mode runs about eight times faster at six times the price, which most genealogists will not need but may enjoy for large transcription batches.

The new Decisions API answers from a fixed set of options in a fraction of a second. It is built for routing and classification, which could suit sorting a pile of unlabeled images into record types, such as deeds, wills, or probate inventories.

Privacy: the item genealogists should watch most closely

OpenAI previewed Private Intelligence, which pairs zero data retention with private safety processing, with private inference coming this fall. Genealogists handle sensitive material, including DNA results, living relatives' details, and unpublished family stories. The hosts called this the most significant item for enterprises and predicted every provider will offer something similar by January 1. If it works as described, it could ease one of the biggest concerns about putting genetic and family data into AI tools. Genealogists should still read the fine print on what is retained, who can access it, and what counts as an approved safety purpose.

Tools for the people who build genealogy software

Sign in with ChatGPT lets users spend their own plan's tokens inside third-party apps. For the small developers behind genealogy tools, that removes the burden of paying for every user's AI costs. Plugin extensions let developers build full applications inside ChatGPT, and the new OpenAI Marketplace gives them a way to reach customers. It is easy to imagine tree builders, record indexes, and DNA analysis tools showing up in that ecosystem.

Cautions

Dots and most of the new products are not yet available in the EU or UK, which leaves many genealogists waiting. The hosts also urged viewers to treat vendor-published benchmarks as marketing, since each lab picks the comparisons that flatter it.

That brings up a term worth knowing: benchmaxxing. It means tuning a model, or the way its results are presented, to score as high as possible on public benchmark tests rather than to be better at real work. A model can post impressive numbers on a coding or math test and still stumble when asked to read a faded deed or untangle two men of the same name. Test any new tool on your own records and your own research questions before trusting it. Whatever an assistant produces, a genealogist's standards do not change. Cite the source, check the original, and keep the human in charge of the conclusions.

The takeaway

Most of what OpenAI shared is about making AI easier to use and easier to fold into daily work. For genealogists, the promise is an assistant that handles the tedious parts of research while you handle the evidence. The new privacy options and cheaper cached input make that more realistic, and the shared Spaces make it easier to work as a team.

Compiled by Katherine and Claude (Sonnet 5.5 Medium). Source: The Neuron's OpenAI DevDay 2026 live watch party on YouTube.

Match the Model to the Task: A Genealogist’s Guide to Saving Time and Tokens

I tokenmaxxed today. I was working with frontier model ChatGPT's Astra on Max level to create a map. The third generation of the map was what I wanted. I asked Astra two more questions and hit my token limit. The questions were simple ones so I should have dropped to a lower model and strength to preserve tokens. Live and learn. The point is that waste comes from paying frontier prices for simple work; or from redoing cheap work that failed. So Claude's Fable 5.1 Medium compiled this article with this human's input and review to help genealogists figure out how to conserve tokens.

Every major AI company now sells a ladder of models. At the top are frontier models such as Claude Fable 5.1 and GPT-6 Astra. At the bottom are small, fast models such as Claude Haiku 5.5 and GPT-6 Luna. The top of the ladder costs up to 100 times more per token than the bottom. The bottom makes more mistakes. Knowing which rung fits which genealogical task saves both money and hours of rework.

The current ladder

Prices are per million tokens, input and output, as of October 12, 2026.

Tier	Model	Input	Output
Frontier	Claude Fable 5.1	$10.00	$50.00
Frontier	GPT-6 Astra	$10.00	$50.00
Upper middle	Claude Opus 5.5	$4.00	$20.00
Middle	Claude Sonnet 5.5	$2.00	$10.00
Middle	GPT-6.1 Sol	$2.00	$10.00
Middle	Gemini 3.1 Pro	$2.00	$12.00
Middle	Gemini 3.8 Flash	$0.75	$3.75
Economy	Claude Haiku 5.5	$0.10	$0.50
Economy	GPT-6 Luna	$0.10	$0.50

Sources for the table are the three companies' price lists. [1][2][3] The Gemini 3.8 Flash price is promotional and doubles on January 1, 2027. [3] Google announced Gemini 4 Argon on September 30 at an introductory $2 and $10, but it was limited to testers and top-tier subscribers at announcement and is not yet on Google's public price list. [3][4]

Where the tokens go

Most genealogists pay a flat subscription, so tokens show up as usage limits instead of a bill. Either way, three things drain them.

The first is price. A frontier model burns through an allowance far faster than a middle model doing the same job.

The second is thinking. Reasoning tokens are billed but never shown. First Page Sage found they made up 22.4% of billed tokens across its 14-task study and were the largest hidden cost. [7] Effort settings matter here: at maximum effort, Claude Opus 5.5 used about 119,000 output tokens per benchmark task, Fable 5.1 about 78,000, and GPT-6 Astra about 27,000. [5]

The third is failure. Cheap models are not always frugal. GPT-6 Luna used about 51,000 output tokens per task against 31,000 for the larger GPT-6 Sol. [6] In the First Page Sage content test, the two cheapest models had 45% and 48% of their drafts thrown away, against 9% for the best model. [7] A discarded answer costs tokens and your time.

Transcribing English handwriting: middle tier, thinking turned down

This is the best-tested genealogy task, and the evidence says not to pay for the top rung. Mark Humphries measured Gemini 3 Pro at a 1.67% character error rate on 18th and 19th century English documents, with the best results at the lowest thinking setting and a cost of about one cent per page. [8] Nicole Elder Dyer found Gemini 3 Flash slightly better than Pro, and on one North Carolina deed it scored 1.22% against 5.61% for Transkribus. She advises turning extended thinking off, because reasoning modes tend to "correct" old spellings. [10] A Dutch test found that limiting the Flash model's reasoning budget raised its score from 47% to 67%. [11]

For transcription, then, extra thinking costs more and can make the result worse. Using Humphries's token counts, my rough estimate is a third of a cent per page at Gemini 3.8 Flash prices and about four cents at frontier prices, before any reasoning tokens. One caution: these tests used earlier Gemini versions, and I found no published handwriting test of Gemini 3.8 Flash. Claude Opus 5.5 can no longer run with thinking switched off, so set its effort to low instead. [19]

Checking a transcription: a second middle model

Spend the savings on verification. Humphries had two other models compare their readings with the first. They flagged 374 disagreements, 106 of them real errors, and a human had to review only about 4% of the text. [9] Names, dates, and places remain the weakest spots, so check each one against the image. [10][14]

Difficult hands and other languages: step up, and still verify

On the University of Basel's benchmark, updated October 6, the top scores for 15th century German manuscripts were Claude Fable 5 at 88.4 and GPT-6 Astra at 87.1. Claude Sonnet 5.5, at one fifth the price, scored 85.9. [12] A 29-language benchmark found the best models still had handwritten character error rates near 30%. [13] Start in the middle, move to a frontier model for pages that defeat it, and use a specialized tool such as Transkribus alongside. [10]

Tables, registers, and index cards: middle tier

Extraction from structured records does not reward the most expensive models. On the Basel benchmark, Gemini 3.8 Flash led on meeting minutes at 95.5, and Claude Sonnet 5.5 led on library cards at 91.4. [12] In the Dutch test, Gemini read a 23-column register in one pass. [11] Dyer notes that censuses and tax lists are still harder than prose. [10]

Economy tier: text in, text out

Use the cheapest models for reformatting a clean transcription, sorting a list, or summarizing typed text. Do not use them to read images. Claude Haiku 4.5 scored 19% in the Dutch handwriting test, against 75% for the best model. [11] Haiku 5.5 and GPT-6 Luna have no published handwriting results.

Writing: middle to upper middle

The First Page Sage data covers the previous generation, but the pattern is useful: Claude Opus 5 and Sonnet 5 needed the fewest revisions, and Sonnet 5 finished cheaper than a model with a lower first-draft price. [7] Denyse Allen reports the same preference for Claude in family narratives. [16] Their successors, Opus 5.5 and Sonnet 5.5, are the reasonable picks.

Evidence analysis and proof arguments: where frontier models earn their price

Correlating conflicting records across a large file is the task where a wrong answer costs the most. Opus 5.5 currently holds the highest Artificial Analysis Intelligence Index score at 58, and Fable 5.1 and Astra sit in the same class. [5] No published benchmark tests genealogical reasoning, so this recommendation rests on general measures. A sensible rule from developer practice applies: start with the cheapest model likely to succeed, and when its answer fails your check, move up a tier instead of retrying. [15]

Finding records: not a chatbot task

Chat models do not search archives. FamilySearch Full-Text Search covers nearly 2 billion records in English, Spanish, and Portuguese. [17]

Quick reference

Task	Tier to use	Effort setting
English handwriting	Middle (Gemini Flash or Pro)	Low or off
Checking a transcription	Second middle model	Low
Hard hands, other languages	Middle, then frontier	Medium
Tables and registers	Middle	Low to medium
Reformatting clean text	Economy	Low
Narrative writing	Middle to upper middle	Medium
Evidence analysis	Frontier	High

Whatever the tier, the Coalition for Responsible AI in Genealogy asks us to verify AI output against the records and to disclose its use. [18]

Sources

1. Anthropic, "Pricing," Claude Platform Docs, accessed October 12, 2026. https://platform.claude.com/docs/en/about-claude/pricing
2. OpenAI, "Pricing," API documentation, accessed October 12, 2026. https://developers.openai.com/api/docs/pricing
3. Google, "Gemini Developer API pricing," accessed October 12, 2026. https://ai.google.dev/gemini-api/docs/pricing
4. "Google announces Gemini 4 Argon as its new frontier model," 9to5Google, September 30, 2026. https://9to5google.com/2026/09/30/gemini-4-argon-announcement/
5. Artificial Analysis, analysis of Claude Opus 5.5, September 22, 2026. https://artificialanalysis.ai/articles/claude-opus-5-5
6. Artificial Analysis, analysis of GPT-6 Sol and Luna, September 22, 2026. https://artificialanalysis.ai/articles/gpt-6-sol-and-luna-push-the-cost-efficiency-frontier
7. Evan Bailyn, "AI Token Usage Cost by Model: OpenAI vs. Anthropic vs. Google," First Page Sage, September 9, 2026. https://firstpagesage.com/seo-blog/ai-token-usage-cost-by-model/
8. Mark Humphries, "Gemini 3 Solves Handwriting Recognition and it's a Bitter Lesson," Generative History, November 25, 2025. https://generativehistory.substack.com/p/gemini-3-solves-handwriting-recognition
9. Mark Humphries, "When Models Disagree...Transcription Accuracy Improves Significantly," Generative History, May 27, 2026. https://generativehistory.substack.com/p/when-models-disagree-transcription
10. Nicole Elder Dyer, "AI-Powered Transcription of Handwritten Documents," RootsTech 2026 syllabus. https://cms-z-assets.familysearch.org/3e/69/1106b1c7ad61708fa9c63d714652/syllabus-for-ai-powered-transcription-of-handwritten-documents.pdf
11. Lieuwe Jongsma, "Can a general LLM really just read handwriting?" Neerlandistiek, June 21, 2026. https://neerlandistiek.nl/2026/06/can-a-general-llm-really-just-read-handwriting/
12. University of Basel, RISE Humanities Data Benchmark, version 0.6.1, updated October 6, 2026. https://rise-services.rise.unibas.ch/benchmarks/p/benchmark-results/
13. Mélodie Boillet, Solène Tarride, and Christopher Kermorvant, "METATR: A Multilingual, Evolving Benchmark for Automatic Text Recognition," arXiv, May 26, 2026. https://arxiv.org/abs/2605.26712
14. Steve Little, "When the Machine Finally Learned to Read: Gemini 3 and the Question of 'Good Enough,'" AI Genealogy Insights, December 16, 2025. https://aigenealogyinsights.com/2025/12/16/when-the-machine-finally-learned-to-read-gemini-3-and-the-question-of-good-enough/
15. Satvik Paramkusam, "AI Model Routing in 2026: When to Use Fable, Astra, Gemini or Muse," Build Fast with AI, September 9, 2026. https://blog.buildfastwithai.com/ai-model-routing-2026-fable-astra-gemini-muse
16. Denyse Allen, "My practical AI toolkit for family history for 2026," Chronicle Makers, December 3, 2025. https://denyseallen.substack.com/p/practical-ai-toolkit-genealogy-2026-chatgpt-claude-gemini-perplexity
17. FamilySearch, "Full-Text Search: A Powerful Tool for Making Genealogy Discoveries," October 18, 2025. https://www.familysearch.org/en/blog/what-is-full-text-search
18. Coalition for Responsible AI in Genealogy, "Guiding Principles for Responsible AI in Genealogy," adopted by the National Genealogical Society, May 2025. https://www.ngsgenealogy.org/wp-content/uploads/2026/01/Responsible-Use-of-AI.pdf
19. Anthropic, "Claude Opus 5.5," September 22, 2026. https://www.anthropic.com/claude-opus-5-5