Introduction
I build kotonia, a character chat service, on my own. Its local LLM runs on two GPUs at home. The chat plan costs ¥500 a month as of November 12, 2026. That plan offers unlimited local-model character chat and voice conversations; image generation and avatars have separate allowances.
Since I already own the GPUs, the marginal cost of local inference is mostly electricity. Hardware purchases, maintenance, and my own time are separate costs. Even so, the setup gives me different pricing options from paying an external API for every request.
If I could offer something good at a low price, surely more people would come. I believed that for a long time.
An easy belief for an engineer
Some markets do reward price comparisons. With APIs and GPUs, buyers compare specifications and costs, and a lower price can be a strong advantage. If that's the kind of buying you're used to, it's tempting to believe that making something good and inexpensive will be enough.
In my service, users were stopping earlier.
The conversation ends before the value can be tested
I followed 1,254 people who registered between August 28 and October 28, 2026, excluding repeat registrations from the same device and excluded or suspended accounts.
| Stage | People | Share of signups |
|---|---|---|
| Sent at least one turn | 632 | 50% |
| Reached ten turns | 162 | 13% |
| Paid | 8 | 0.64% |
Half never sent a message. Only 13% reached ten turns.
That 13% is not the share who saw a pricing screen. This analysis didn't directly measure pricing exposure. Some users may have read the pricing page before registering, so it would be wrong to say everyone else never saw the price.
What it does show is that many people leave before trying the benefit of a long conversation without counting each turn. Explaining long-term savings is unlikely to do much for someone who hasn't managed to send a first message.
It doesn't prove that price dissatisfaction wasn't a cause either. It tells me there was an entrance to investigate before changing the price.
Before comparison comes discovery
When I played competing services, the first thing I chose was a scenario, presented through an image and a title. Comparing models or prices came after thinking, “I'd like to try this story.”
My service sent new users straight into a conversation with a default character. That character accounted for 75% of first conversations, and only 19% of users had visited the catalog. Switching partners inside the conversation screen wasn't counted as a catalog visit.
There was little room to find a story they wanted before we started competing on quality or price.
People who visited the catalog did tend to talk longer. But more engaged people are also more likely to open a catalog, so that correlation doesn't establish an effect. We decided to test it by assigning different destinations after signup.
Turn low marginal cost into a useful interaction
I don't want to throw away the pricing advantage. The problem was explaining it only through a price list.
Some people have had more than 1,000 turns with a particular character. A flat-rate model suits that use. We haven't asked them whether low price is why they stay, though.
To make it a reason to arrive, describe the experience. Alongside “inexpensive,” say what it enables: continuing a conversation without keeping count of every turn.
Recently, I added a button that drafts the user's next reply. Reusing the conversation cache avoids reevaluating most of the history. I explain the mechanism in the cache investigation.
Reply drafts aren't charged individually. They aren't unlimited calls, either. One generation per conversation turn produces five drafts, which the UI lets users browse. An abuse-prevention limit also caps requests at 120 per hour. Generating the output still uses compute, and occupying a slot during busy periods still has a cost.
If we can keep marginal costs low, we can spend that advantage on helping a user find their first words. Make it useful at the entrance, as well as making a price comparison.
The order of work
- Help visitors find a story they want to try. Make the catalog an entrance and let card headings describe the story, rather than only naming the character.
- Help them send a first message, with opening reply options and generated drafts.
- Explain at the entrance which conversations they can continue without counting turns.
- Then build acquisition routes that reach people who care about price.
It's tempting to start at step four. But if many people still get stuck at steps one and two, new arrivals may stumble in the same place.
As of November 12, 2026, we haven't measured retention or payment rates after the changes. Better onboarding leading to more revenue remains a hypothesis to test.
The work between making and being chosen
Between making something good at a low price and getting chosen, there was another job: bringing that advantage to a place where users could experience it.
As the person building it, I had underestimated that job.
Even my impression that our model quality compares well with competitors is still just one person's opinion. Low price, reasons to choose a service, and reasons to stay need to be measured separately.
