Every few weeks in 2025, a new model came out. GPT-5. Claude Opus 4.1. Gemini 2.5. GPT-5.1. Gemini 3.0. Claude Opus 4.5. GPT-5.2.
Each one was supposedly better than the last. Better at reasoning. Better at coding. Better at context understanding.
But here’s what I realized. Companies can’t actually take advantage of this speed.
Let’s say you’re a company that deployed GPT-4 in your application six months ago. You tuned it. You got it working. Your system is stable. Your customers are happy.
Now GPT-5 comes out. Should you upgrade?
Theoretically, yes. It’s better.
But practically? Your engineering team has to re-test everything. They have to make sure the new model doesn’t break existing functionality. They might have to re-tune prompts. They have to check cost implications. They have to plan a rollout.
That’s 2-4 weeks of work for a team of 2-3 engineers.
But a new model comes out in 2-3 weeks.
So you’re perpetually behind. You’re using last-quarter’s model because you haven’t finished upgrading to last-month’s model.
This is creating an interesting dynamic. Model improvement speed is outpacing company adoption speed.
Which means the benefit of having the latest model is… not that big. Because you’re probably not using the latest model.
I think companies will start asking: “Do we actually need to upgrade every six weeks? Or can we stay on a stable model for a year or two?”
If the answer is “stay stable,” then it doesn’t matter that models are improving fast. It matters that models are stable. The advantage goes to whoever can give you a rock-solid model that doesn’t change.
Right now everyone is focused on “latest and greatest.” But I think the market will shift to “boring and reliable.”
OpenAI might win because they have the ecosystem lock-in. But Google or Anthropic could win if they can offer “this model, same features, supported for 24 months, no surprises.”