The statement that makes you think
In early April 2026, Brad Lightcap, OpenAI's chief operating officer, dropped a line that went viral: "By the end of the year, we'll laugh at today's models. They'll seem ordinary."
To put things in perspective: he was talking about GPT-5.4, a model released just days earlier, which was already handling 5 trillion tokens a day and which a good share of the world's developers now consider the absolute state of the art. And according to its own creator, it will be obsolete before Christmas.
It isn't just OpenAI that's speeding up
What makes this statement credible is that it doesn't concern a single company. In December 2025, Anthropic, Google and OpenAI each released major updates to their flagship models in the same month, something unprecedented in the industry's history. Since then, the pace hasn't slowed.
OpenAI went from GPT-5.1 to GPT-5.4 in under four weeks. Anthropic released Claude Opus 4.7 two and a half months after Opus 4.6. Google rolled out its Gemini 3.1 variants on a weekly cadence. And meanwhile, DeepSeek proved that you can train a model rivalling the best for just $294,000, where GPT-5 cost hundreds of millions.
According to the consultancy Deloitte, the cost of making an AI "think" has been divided by 280 in two years. What cost 280 euros in 2024 costs 1 euro today. And yet overall spending on AI is rising, because usage is growing even faster than prices are falling.
Why is it getting faster and faster?
The reason is simple, almost dizzying: the labs now use their own AI to build the next versions. AI improves because it helps itself progress. Each new version speeds up the creation of the next. This is what Brad Lightcap calls "the collapse of training cycles".
In 1965, an engineer named Gordon Moore observed that computing power doubled every 18 months. That rule shaped 50 years of innovation. What we're seeing with AI models in 2026 is a far more compressed version of that phenomenon: doubling no longer takes 18 months, sometimes a few weeks are enough.
What it changes in practice
For an individual, it's rather good news: the tools improve without the price going up. For a business, it's more complicated. Betting on a specific model today means accepting that it will be superseded before the end of the year. The real skill in 2026 is no longer choosing the right model, it's being able to switch quickly when the next one arrives.
Some developers have already learned the lesson: instead of building products around a specific model, they build systems capable of plugging in any model within hours. The architecture becomes more important than the AI itself.
The question nobody is asking yet
There is something troubling about this acceleration that product announcements never mention. If a technology is evolving faster than our ability to understand it, and even the engineers building it can no longer predict exactly what a model will do before testing it, are we still really in control?
This is no longer a science-fiction question. It is a very concrete question that the world's largest laboratories are asking themselves right now. Anthropic, for that matter, has built its entire brand image around an answer to this problem. But building guardrails on something we do not fully understand is a bit like designing the brakes of a car whose top speed is still unknown.
It is precisely this debate, and its implications for all of us, that we will explore in greater depth in a forthcoming article.