Dustin Snoap
to you There is a graph somewhere, in a very expensive conference room, that goes up and to the right forever. It has been in that conference room for about
There is a graph somewhere, in a very expensive conference room, that goes up and to the right forever. It has been in that conference room for about three years now. Nobody has told it what an S-curve is, and frankly, at this point it seems awkward to bring it up.
This is more or less the story of AI progress in 2026.
The leap from GPT-3 to GPT-4 was genuinely remarkable. GPT-3 could write a sentence that sounded like a human had written it, provided the human in question never had an original thought. GPT-4 could reason, code, synthesize, and argue, sometimes correctly. People who encountered it for the first time had the specific expression of someone who has just watched a dog play chess. Not well, perhaps. But the dog was playing chess.
Then GPT-5 arrived. And it was fine.
Not bad. Not a disappointment, exactly. Fine in the way that a slightly faster elevator is fine, which is to say you would only notice it if you had spent the preceding months being told the elevator was going to change everything.
The progress is not an illusion. So why did it feel like nothing happened?
Between GPT-3 and GPT-4, OpenAI released one significant intermediate model. Between GPT-4 and GPT-5, the industry released dozens. The gains were real. They just arrived in installments, spread across eighteen months of incremental releases, each one quietly raising the baseline until GPT-5 landed in a market that had already absorbed most of what it had to offer. OpenAI managed to deliver a genuinely large improvement and still disappoint everyone. This is, by any measure, an impressive achievement.
The cost side of the ledger is where it gets interesting.
Training GPT-3 cost an estimated $4.6 million. Training GPT-4 cost somewhere north of $100 million, by Sam Altman's own admission, with some estimates reaching $540 million. Training runs for GPT-5 were reportedly around $500 million each. Billion-dollar training runs are coming. Ten-billion-dollar models are a foreseeable next step.
Every generation costs dramatically more than the last. Every generation delivers improvements that are real but increasingly difficult to perceive without a benchmark and a quiet place to sit.
Every major technology eventually confronts the same inconvenient truth: things cannot go up and to the right forever. The universe has opinions about this and they are, as it turns out, non-negotiable.
There is always the same pattern. An early slow crawl. A breakthrough that unlocks scale. An explosive middle phase where each dollar of investment produces results dramatic enough to generate breathless op-eds about the end of work. And then the upper bend, where the physics and economics arrive uninvited, sit down, and begin explaining how things actually work.
The explosive middle of the AI curve is behind us. The upper bend has arrived, as upper bends tend to do, without sending advance notice or offering a refund.
The question is what comes next.