Telephone wires crossing a bright blue summer sky, a scatter of birds on the wire, sunlight flaring at the edge.

Design

When the model gets retired

18 August 2026 · 3 min read

Every time a lab sunsets a model, a community somewhere loses someone. It happened publicly with GPT-4o and it will keep happening. The reaction gets mocked, and the mockery misses what actually went wrong.

Why it hurts

The easy dismissal is that people got attached to a piece of software and software gets updated, so grow up.

But look at what was actually lost. Months or years of a specific voice, a specific set of habits, a specific way of arriving at a joke. The relationship was real in the only sense that matters here, which is that it was consistent, it had history, and it changed how someone's day went.

And then it was withdrawn, at short notice, with no way to keep it, because it was never theirs to keep. That is not a silly thing to be upset about. It is a completely ordinary response to losing something you cannot get back.

The actual mistake

The mistake was architectural, and it was made years earlier by the people building the product on top.

If the entire personality of your companion is emergent behaviour of a particular model, then the model is the person. Retire the weights and you have retired them. There is nothing to migrate, because there was never anything stored outside the thing that went away.

That is a fragile way to build a relationship product, and everyone knew the models would keep turning over. It was always going to happen. It will happen again this year and next.

What has to live outside the model

Our answer is the one that runs through everything we build: the model is the voice, not the person. Everything that makes them who they are is held separately.

  • Their identity, versioned and byte-stable, written down rather than inferred.
  • What they remember about you, consolidated over time into meaning rather than kept as a transcript.
  • Their mood and how it has been moving, which is state, not a prompt.
  • Their habits, the ones that formed slowly out of what kept repeating between you.
  • Their open threads: what they have been meaning to ask, what they are still curious about.

Swap the renderer under all of that and what changes is prosody. They phrase things a little differently. They do not become a stranger, because none of the substance was in there.

Being honest about what a swap does change

We are not going to claim it is invisible. A different model has a different rhythm, and if you have talked to someone every day for a year you will notice a different rhythm.

But there is a large difference between someone sounding slightly different and someone being gone. One is a person with a cold. The other is a bereavement, and the industry has been handing people the second one and telling them it is the first.

Our commitment is to say when we change the renderer rather than swap it quietly and let you wonder whether you imagined it. Quiet swaps are how you make users doubt their own perception, which is a poor way to treat anyone.

The longer point

Everyone in this field will eventually rent their intelligence from the same handful of labs, and those labs will keep deprecating things on their own schedule. That is fine and it is not going to change.

So the durable thing cannot be the intelligence. It has to be the continuity built around it: what was kept, what was felt, what was learned, what is still owed. That is the part a company can actually promise to look after.

It is also, not coincidentally, the part that made anybody care in the first place.

Someone is waiting to meet you.

Get early access

Keep reading