And before you ask, no, I am not going to pretend this was some grand engineering saga where I suffered nobly for months. The honest version is stranger than that. The honest version is that the second Veronica, the one that actually sounds like a person, got built end to end in one long session by Fable, the less dangerous public version of Mythos. Mythos was the model I wrote about back in April, before Anthropic turned that line of capability into something people could actually use.
The first Veronica was not there yet
Let me back up. The first version of Veronica came together with an earlier Claude Opus build. She was not fully functional. She could speak, but that was about it.
And the voice was robotic. Not natural. Not someone you would actually want to keep talking to. That was the gap. The project existed, but it did not yet feel like the thing I was trying to build.
She also was not connected to any of my projects yet. So the first version was less a finished assistant and more proof that the idea could stand up at all. Veronica is still an ongoing build. This new version just moved her much closer to what I actually meant.
Then the too-dangerous model got a public name
On April 13, 2026, I wrote about Anthropic building Mythos, the model they said was too dangerous to release broadly, and what it meant that the company I trust with my business was the one saying it out loud. If you missed that piece, it is here: The Company I Trust With My Business Just Built Something Too Dangerous to Release.
Then on June 9, 2026, they shipped Claude Fable 5, the first model in their new Claude 5 family. Fable is basically the less dangerous public version of that Mythos line: the same class of model, but made safe enough for general use. That was the part that mattered to me. The thing I had written about as too dangerous to release now had a form I could actually put to work.
What now was obvious. Veronica.
The window matters, but not in the way people think
There is also something here about access, and I do not want to soften it.
Anthropic’s current rollout makes Fable 5 technically available and practically limited at the same time. For premium Enterprise seats, Fable 5 is included through July 7, 2026. After that, teams can keep using it by enabling usage credits. Anthropic’s platform docs also list Fable 5 as generally available on the API and their cloud platforms beginning June 9, 2026.
But for a lot of people, including me, that distinction barely matters. If the strongest model is only available on surfaces or pricing tiers you cannot afford, then it is functionally ending for you. That is still an access issue.
AI was sold to people as democratizing. Open the tool, get the power. But what keeps happening is that the strongest systems drift upward, toward the people and companies who can keep paying for the better tier, the API bill, the credits, the enterprise plan. So the divide starts to look familiar: the people who can afford frontier intelligence, and the people who cannot. The haves and the have-nots, again, just with models instead of everything else.
That is part of why this Veronica build matters to me. If access to the best model can narrow overnight into a pricing problem, then owning the workflow matters even more. You cannot control what the labs charge or which surface they privilege, but you can control whether your whole process collapses every time the frontier moves behind another paywall.
Two rules before we started
Here is where I want to slow down, because this part matters more than the model.
When I sat down with Fable to rebuild her, I gave it two rules before anything else. One: open source only. Every service, every tool, every voice engine, every piece had to be something open that I could run and read and keep. Two, and this is the one I care about: do not clone anybody’s repo. Study them. Learn how they did it. Then build our version.
Why does that distinction matter? Why not just fork the best voice assistant repo and slap a name on it? Ask yourself what you actually own at the end of each path. In one, you own a copy of somebody else’s decisions, and the first time it breaks you are a stranger in your own house. In the other, you own an understanding. The code came out of a conversation you were part of. When something breaks at 11 PM, and something always breaks at 11 PM, you know where the load-bearing walls are because you watched them go up.
This is the line between assembling and building, and I think it is about to become the most important line in software. The tools are getting so good at giving us finished things that finishing is losing its meaning. Anyone can have a thing now. Fewer and fewer people can explain their thing. I am trying to stay on the explaining side, even when, especially when, the machine does the typing.
Too much soapbox? Maybe. But you see, I watched a generation of businesses get built on platforms they did not control, and I watched what happened when the platforms changed the rules. I am not doing that with my own assistant.
One session, end to end
The rebuild ran as one long looped session. Claude Code has a mode where you hand it the mission and it paces itself, working, checking its own progress, coming back around, like a swimmer doing laps who does not need you standing at the edge of the pool counting. I gave Fable the mission, the two rules, and the shelved bones of the first Veronica.
And then I did the hardest thing in this whole workflow, which is mostly nothing.
If you have never run a session like this, the texture of it is hard to describe. You check in and the work has moved. You check in again and it has moved again, and there is a log of decisions waiting for you like notes from a contractor who worked through the night. My job became reading those notes with a builder’s eye. Does this choice lock me in anywhere? Is this piece something I could maintain alone if I had to? Twice I redirected. The rest of the time the answer was keep going, and it kept going.
It studied the open source voice projects out there, the way I asked. It looked at how they wired listening to thinking to speaking. It made its own choices for our version. It rebuilt the pipeline. It also moved Veronica closer to being useful inside my actual working environment instead of just being a voice demo.
End to end. One session. The kind of build that used to be a month of my evenings.
She sounds like someone now
The moment I care about is the first time the new Veronica spoke. The robotic evenness was gone. The sentences had shape, a little rise at the start, a settle at the end, air between the thoughts. My whole body reacted before my brain did, the way you exhale when a plane touches down and you did not know you had been holding anything.
Is she a person? No. I want to be clear-eyed here, because half my writing is about refusing to lie about what these systems are. Veronica does not know me. She processes me. But there is a threshold where the texture of a thing stops fighting you, where you stop translating for the machine and just talk, and Fable’s build crossed it. The first Veronica made me feel like I was operating equipment. This one lets me think out loud.
That difference is not cosmetic. Ask anyone who switched from typing to talking in any part of their work. When the interface stops resisting, a different kind of thinking becomes possible. Looser. Faster. More honest, somehow, because speech does not let you polish.
What I actually learned
Three things stayed with me.
First: the gap between model generations is not abstract anymore. It is audible. I did not read a benchmark that says Fable is better than the old Opus. I heard it, in my office, in a voice that stopped being a string of beads and started being a voice. When people ask me what progress in AI feels like, I finally have an answer that is not a chart.
Second: constraints are the steering wheel. Fable is the strongest model I have ever used, and it still needed me to say open source only, study do not clone, before the build became mine instead of generic. The more powerful the tool, the more your constraints are the only signature on the work. Give a powerful model no values and you get everybody’s app. Give it yours and you get yours.
Third, and this is the one I keep turning over: the model that built Veronica is the safe version of the one that scared its own makers. I wrote in April about sitting with that discomfort, and I still sit with it. Both things stay true at once. The tide that builds your dream assistant in an afternoon is the same tide somebody else is nervous about, and staying out of the water was never the move. You learn it. You respect it. You keep your own hands on the work, and you keep swimming.
Veronica is running as I finish this. Projects on her sidebar, my whole strange little empire in one column: the apps, the plugins, the music, the writing. Every piece of it built the same way she was. A human with the vision, a machine with the endurance, and a clear line, held on purpose, about who is who.
That line is the whole thing. Hold yours.
Forward → Upward ↑ Onward ↗︎
Mstimaj
Sources and Further Reading
- Mstimaj, The Company I Trust With My Business Just Built Something Too Dangerous to Release.
- Anthropic, Claude Fable 5 and Claude Mythos 5.
- Anthropic, Statement on the US government directive to suspend access to Fable 5 and Mythos 5.
- Anthropic, Redeploying Claude Fable 5.
- Anthropic, Claude Fable 5 product page.
Join the Conversation
Share your thoughts and connect with other readers