Fable 5: The Night Agent

Three days with a frontier model, and a glimpse of how much bigger we should be thinking.

A note as this goes out (the evening of June 30, 2026). I set out to write a postmortem. As I was finishing, Anthropic announced that the Department of Commerce had lifted the export controls, and that access would begin returning the next day. So read this less as an elegy for something lost than as a field guide for something arriving; the window I describe is reopening now.

For three days in mid-June 2026, we had access to a new frontier AI model that gave us a glimpse of the future, and perhaps changed the way we get access to the best models.

The model was Anthropic’s Claude Fable 5. I liked it enough to burn through my usage limits and pay for more; I was planning to keep paying. What I did with it in those three days convinced me of something I want to put plainly, up front: most of us, myself included, are still thinking too small about what these tools are for.

The moment the engine changed. Midway through a long project, I swapped the engine. This is the message where I told the assistant it was now running on Claude Fable 5, released that morning, and its reply. It confirmed the change, thanked me for marking it, and then, without a pause, laid out exactly where we stood: the version history in order, the decisions I had locked, the gate we had just cleared, the one step still waiting on my go, even how it coordinates with a second AI collaborator through shared files. No re-briefing, no lost thread; the new model simply picked up a messy, real project and held all of it at once. That instant command is what the next three days were made of.

So let me show you, smallest to largest, what one researcher built; then I will tell you why I think the ceiling is so much higher than we treat it.

What I wanted to know

My test was not whether Fable 5 was clever. It was this: how closely could it keep sound genealogical standards and methods foremost in its processing? Could it work the way we are supposed to work; honoring sources, separating what a record says from what we infer, keeping uncertainty visible, refusing to invent evidence, holding citations and provenance in view, and protecting living people?

Three pieces of work answered, each bigger than the last.

One record

I started where everyone starts: one hard record. I handed Fable a single image, a WWII draft-registration card, and asked it to read the card the way a genealogist would.

That day I posted a first reaction to my genealogy-and-AI group:

First Glance: Claude Fable 5 and Claude Mythos 5. Anthropic released their long-expected Claude 5, Fable and Mythos. Attached and below is a quick first glance at how Claude Fable 5 handles genealogical tasks, the one-shot processing of a record.

What came back was not a confident paragraph about the man on the card; it was the card taken apart into evidence. What the handwriting literally showed. Which details were firsthand and which were hearsay. What each line could and could not support. Where it had to mark uncertainty rather than guess. One image, one pass, and the output was already shaped for review instead of belief. That is the humble, repeatable version of the whole story, and the one any genealogist can try today with the tools already on their desk.

What Fable did with one record. I gave Fable one image, my grandfather's 1942 draft-registration card, and asked it to read the record the way a genealogist would. This is what it returned. Not "here is when he was born," but a careful draft: because a single card carries only secondhand information about birth, it held the conclusion at Probable, not Proved, and said so. It laid out what the card supports, flagged the wife as a likely-but-unconfirmed inference, wrote a full citation with the parts it could not verify left in brackets rather than invented, and named the exact next sources that would raise the finding to Proved. Then it logged its own gaps: the card's reverse unread, the employer line illegible, the registration date inferred. Evidence, shaped for review, not belief.

One place

Then I gave it something bigger: not a record, but a place. I asked it to build a circa-1872 map of Fauquier County, Virginia, paying attention to the area around Catlett; the Virginia community where I have lived for the past twenty years, and a place I know well enough to tell whether the map got it right.

This is the piece I would hand a skeptic, so I will let my own account from that day stand:

Claude 5 Fable can generate maps. But it ain’t cheap or quick... It took Fable about 45 minutes, burned through about 1.3 million tokens, used 23% of my half-day cap and about 8% of my weekly cap.

And the part that matters, how it worked, in its own words that I posted:

Prompt one: “describe a process.” Instead of answering from memory (where it’s easy to get details confidently wrong), Fable launched eight research assistants — separate AI sessions, each with its own assignment. Four searched the live web: digitized Civil War-era maps at the Library of Congress, railroad corporate histories, Virginia county records. The other four acted as skeptics, re-checking every claim the first four made and trying to disprove it. That fact-checking caught real errors — for example, in 1872 the railroad through Catlett was the “Orange, Alexandria & Manassas,” a name even an official government document gets wrong.

Prompt two: “generate the map.” ... Fable downloaded genuine geographic data — the county boundary, every stream, the railroad’s actual path — then wrote a small program to draw the map, including only features the research had verified. Crucially, Fable could look at the resulting image, the way you would. Fable spotted a marker at the wrong road crossing and labels overlapping, fixed the program, and re-drew. Four drafts in, the map was honest enough to ship.

About 45 minutes, mostly machines reading.

Look at what actually happened there. Fable did not draw a pretty map from memory; it ran a researcher’s process. It sent out copies of itself: four assistants to gather sources, and four adversarial fact-checkers whose only job was to attack the first four’s claims and catch their errors. It drew only what the evidence verified, looked at the result the way a person would, found its own mistakes, and fixed them. The map it produced was a first-draft reconstruction, not a finished product; but it was sourced, and that changes the questions I can ask: who were the neighbors, which courthouse mattered, what route connected two families, how land moved through kin and FAN networks. Our ancestors did not live in record sets; they lived in neighborhoods.

The lesson copies even without Fable: ask the model to build the research process, not just summarize the place.

One place. When I asked for a place instead of a record, this is what came back, about forty-five minutes and four drafts later. It is a first edition, version 0.1, and it says so; a starting point, not a finished map. But look at the legend: every line is graded by how well the evidence supports it, documented alignment, schematic, or inferred, so an honest map never claims more certainty than its sources allow. The inset is Catlett's Station in 1872, the corner of the county where I live now, drawn from the archives rather than from memory. The railroad crossing it carries the name the fact-checkers verified, not the one an official document got wrong. The next step is not another prompt; it is the courthouse: deeds, tax lists, plats.

One folder

The ceiling I will describe carefully, because it doubles as a warning. I pointed Fable at a folder of uncatalogued record images, opaque filenames, mixed types, no stated question, and let it work overnight. By morning it had designed and run an extraction process and produced more than a thousand evidence-layered findings. Each was tied to what the handwriting showed; each was classified by source, by information, and by evidence; most carried a reading-confidence level and an alternate reading where the hand was ambiguous; and where a value touched a living or recent person, it left a withholding flag in place of the value.

"Good morning, Sunshine": the night agent's morning report. The night before, I handed Fable a folder of scrambled record images and went to bed. This is what I woke to: not an answer, but a night's honest work, reported like a colleague's. Working as an assembly line, it protected the originals and fingerprinted all 155 images, catching on its own a planted near-duplicate whose birthdate had been blanked, and refusing to be fooled. It sorted them, tested its method on twelve records for zero fabrications, then read the rest in parallel. Partway through, my account hit its monthly spending limit; the run did not crash. It wrote an incident report, saved its work, and stopped clean. What it could not finish, the quality checks and a hundred records still waiting, it left for later, and said so.

It did not finish the folder, and the planned quality-assurance pass never ran. I am telling you that on purpose: this was a demonstration, not a finished benchmark. But it showed the shape of the thing. And one moment showed why a person still has to check for privacy after the AI has done its part. Among the records was a near-duplicate pair, where one copy had a field that looked deliberately blanked by a human hand. A careless system fills that blank from the twin. Fable noticed the difference, refused to infer the hidden value, and set the pair aside for a human to judge. Treat a redaction as a decision someone made, not a gap to be closed: that is the lesson worth publishing. The private content stays private.

The stronger the extraction, the more serious the privacy responsibility becomes. Frontier models can now help build first-pass extraction systems for messy collections of records; and that capability raises the bar for privacy, it does not lower it.

The pattern, and a word about “gates”

A record, a place, and a folder look like three different jobs; they share one shape. Fable was strongest treated as a builder of research machinery, not as an oracle. It made ledgers, scripts, schemas, checklists, and gates; work that was useful precisely because it could be inspected.

A word on gates, because the idea is unfamiliar to most of us and genuinely useful. A gate is simply a checkpoint built into a process: the work stops there and cannot move forward until something is satisfied, a test passes, or a person approves. Picture a gate on a fence; nothing gets through until someone opens it. When I had Fable rebuild my Genealogical Research Assistant, the reusable system I use to teach, it did not just write a better prompt; it built the thing with tests and review gates, so that nothing could “ship” until it had passed its checks. That is what separates a toy from a tool: standards, tests, boundaries, and a defined way to behave when something fails.

The night shift. At eleven at night I asked it to propose a plan it could run overnight, on its own. It did not just propose one; it ran it. Five blocks of work, ten agents in parallel, a backup made before it changed anything, its own output cross-examined by verifiers who checked 180 claims and caught six small errors before a single file was finalized. Everything came out verified and lint-clean. Then it handed me a morning state: what it had finished, what was still waiting on my judgment, what needed daylight. It even closed the way a good colleague would, telling me to sleep well because the desk was ready. This is the night shift. Give it the goal, the guardrails, and the gates, and it works while you don't.

So the practical question is not whether to trust an AI to “do genealogy”; we should not hand over judgment that way. The better question is whether AI can help us build better scaffolding around the work: better source control, better uncertainty capture, better research plans, better privacy gates, better first-pass extraction. In this short window, the answer was yes.

What you can copy now

Most genealogists will not have Fable today. That is not the point; copy the shape of the work.

• Ask for workflows, not just answers. Instead of “Who were my ancestor’s parents?”, ask for a source-led research plan, a locality-map workflow, a conflict table, or a citation checklist.

• Require uncertainty. Make the model mark what it reads confidently, what it reads only tentatively, and what it cannot fill without guessing.

• Separate extraction from conclusion. A record can say many things before it proves anything; ask for source, information, and evidence layers first, and decide what a claim means later.

• Keep raw AI output private until a person reviews it, especially with records that touch living or recent people.

• Save your failures. This is the habit I keep recommending: when you find a task today’s AI almost does--tantalizingly close but not right--save it. When a stronger model arrives, the first thing to try is the pile of near-misses. Much of what amazed me in June was last winter’s failures, retried.

Where Fable 5 stands now

Updated the evening of June 30, 2026.

The story turned tonight. Fable 5 had been offline since June 12, when a government export-control directive pulled it and the more restricted Mythos 5 from service. This evening, Anthropic said the Department of Commerce had lifted those controls and that it would begin restoring access the next day. What that means for you is plain: the tool in these pages is not a memory. It is about to be back.

A word on cost and access

When Fable was briefly here in June, monthly subscribers got only a small allowance that ran out fast, and paid access was not cheap: about $10 per million input tokens and $50 per million output, where a million tokens is roughly five and a half Harry Potter books. Two lessons follow. First, watch your usage closely; the meter moves quickly. Second, spend that expensive intelligence where it earns its keep: strategic planning, ambiguous judgment, and supervising other models, letting Fable hand the routine work down to cheaper ones like Opus and Sonnet. And know that when access returns, subscribers may again get only a limited taste before per-token credits, purchased in advance, are the price of going further.

The bottom line

Here is what I most want you to take from three days with a frontier model: most of us are not ready for the power that is about to be available to us again, and we are thinking too small.

We treat these tools as task-doers; transcribe this, summarize that, answer that question. That is where everyone should start, and you should master it. But it is the floor, not the ceiling. Once you are comfortable, start thinking in larger units: not single tasks, but workflows; not single questions, but research processes; not “help me with this record,” but “help me build the project.” A sourced county map. A reusable research assistant with its own tests and gates. An overnight first pass across a whole folder, with uncertainty and privacy built into the output.

Fable was around long enough to give a glimpse, not long enough to become routine. In that glimpse I watched a careful researcher’s process; gather the sources, cross-examine them, draw only what is verified, look at the result, fix it, keep uncertainty and privacy visible; run mostly by machines, mostly while I watched. It did not remove the need for expertise; it made expertise more levered.

So save your failures, and start deciding what you would build. The next window may be short; but it is coming, and it will reward the genealogists who already know how big to think.

Postscript

I wrote the paragraph above believing the next window was coming. I did not expect it that same evening. A few hours later, Anthropic said the export controls were lifted and access would return the next day. So do not wait to decide what you would build. The window is already reopening.

The window reopens. Anthropic's announcement, the evening of June 30, 2026: the Department of Commerce had lifted the export controls, and access would begin returning the next day. The postmortem became a preview while I was still writing it.

Sources

• Anthropic, “Claude Fable 5 and Claude Mythos 5,” June 9, 2026.

• Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5,” June 12, 2026.

• Anthropic (@AnthropicAI), post on X, June 30, 2026: the Department of Commerce lifted export controls on Fable 5 and Mythos 5; access to begin restoring the next day.

• Stephen Little, “First Glance: Claude Fable 5 and Claude Mythos 5,” Genealogy and Artificial Intelligence (AI) group, Facebook, June 9, 2026. https://www.facebook.com/groups/genealogyandai/posts/2064698467472834/

• Stephen Little, “Claude 5 Fable can generate maps. But it ain’t cheap or quick,” Genealogy and Artificial Intelligence (AI) group, Facebook, June 9, 2026. https://www.facebook.com/groups/genealogyandai/posts/2064825660793448/

---