Today (on a Sunday) I took our own AI models off our own app and replaced them with a competitor's model.
This is the story of how the decision happened in 24 hours, what it was like in the 4 days leading up to it, what it cost me to see it, and what I think other founders can take from it.
Day 1 - my own product told me the truth
We're an AI model company. Or we were.
We spent months training small open-weight models for honesty - models that say 'I don't know' instead of inventing answers.
I believed in them. I still do, for what they are.
Then I built our coding CLI and put our models inside it.
They failed. Our 4B failed. Our 9B failed. Multi-step agentic work, the thing our flagship product needs - they couldn't hold it.
The worst part? Our own model card predicted it. We'd published 'not recommended for autonomous agentic loops' ourselves, weeks earlier.
I knew it on paper. Watching it happen is different.
Day 2 - I noticed my own behavior
That same week, a new frontier model came out.
Faster and better than the one I'd been using every day.
I switched my daily coding agent in one afternoon.
No hesitation. No loyalty. The old one was great last month - didn't matter.
Then it hit me: I am the market.
I switched in an afternoon despite having reasons to stay loyal. That tells me about my own switching threshold; it does not show that every user or enterprise behaves the same way.
I was surprised by how little loyalty I felt to my own daily tool.
Whatever we built had to survive that fact, not deny it.
Day 3 - the spreadsheet
I looked up what the best open model in the world costs to use through American inference providers.
About $1 for a serious agentic coding task. Sometimes half that.
I'd loaded $100 into an API account on July 1 to build our CLI.
12 days of heavy testing later - full workloads, everything - I'd spent $31.
Meanwhile, we were paying about $5,000 a month per GPU to serve our own small models. To a handful of users. Models a laptop could run.
I sat with that for a while.
We weren't serving our models because it was right for users.
We were serving them because they were ours.
There's a word for that, and it isn't strategy.
Day 4 - the call
This morning (again, a Sunday), I called Ayush, our engineer.
I laid it out: take our models off the GPUs, serve the best open model instead, cap our markup at 15%, publish everything - the costs, the prices, what's behind every product name.
Editorial update, October 11, 2026: The 15% cap, hosted-model replacement and publication commitments describe the July 12 decision. They are not a current pricing guarantee. The dated product announcement and thirty-day postmortem provide the follow-through described in this archive. The underlying task outputs, acceptance criteria, rejected-output counts and measured review times are not supplied, so this is a decision account rather than an audited comparative trial.
Keep training our own models.
But set one rule: they only go back into the product by beating the best alternative in public, on published evals. If ours loses, we tell people to use the winner.
He agreed in 20 minutes. When the strategy is honest, the meetings get short.
What I learned
Your own usage is the tell.
I could have read 100 market reports about model switching.
Watching myself dump my daily tool in an afternoon taught me more.
If you want to know where your market is going, watch what you do when nobody's watching.
Sunk cost arrives dressed as mission.
'We serve our own models' sounded like conviction.
It was $5,000 a month of not wanting to admit the benchmarks applied to us.
The more noble the reason sounds, the harder you should check it.
Our model was not the moat I hoped it would be.
Editorial update, October 11, 2026: This section was originally headed "Nobody has a moat, and that's information, not doom.". Its narrower heading preserves the original section link here.
Some users, including me, switch quickly. Open weights create more serving options, but deployment, support, safety and reliability still cost money.
The lab that made the best open model can't even win at serving it - American providers undercut them on their own weights.
For us, the lesson was to stop relying on a model advantage we had not demonstrated and invest in the practices we could sustain.
For us that's trust, cost discipline, and the ability to change our minds in 4 days.
Speed of surrender is an asset.
The pivot took 24 hours because we're 3 people with no investors to convince and no sunk narrative to protect.
Our small team and decision authority helped this call move quickly. I cannot infer how every larger company would respond.
Being small helped us make this particular decision quickly.
Where that leaves us
We're tiny. A handful of users. A few hundred model downloads.
This post will be read by more people than use our product.
I'm fine saying that, because the entire company is now a bet on one thing: building a track record of telling the truth, after seeing how readily I switched and how weak our own model advantage proved.
Cheapest possible prices, capped markup, honest meters, public evals - and when we get something wrong, you'll read it here first.
The full product announcement - what changed, the pricing, all of it - is on the SimpleDirect blog.
The CLI ships this week. I'll publish the numbers, including the bad ones.
Editorial update, October 11, 2026: That was the July launch plan. The later coding-agent account links the public repository and reports the subsequent beta limitations.