On August 7, 2025, GPT-5 was released as the new frontier model, delivering state-of-the-art results across the most relevant benchmarks. It feels like that happened yesterday. And in a very real sense, it was practically yesterday.
Now, less than a year later, GPT-5.6 Sol crushes that frontier model across every benchmark that still allows a direct comparison. Even those tests barely communicate the scale of the progress or the gap between models now arriving almost every month. At the same time, they keep getting cheaper.
When benchmarks are no longer enough
At some point, benchmark comparisons stop being useful. The scores rise and the test names change, but a chart cannot fully explain what a model is now capable of doing.
Recently, an internal OpenAI model independently solved an almost 80-year-old Erdős problem concerning unit distances. External mathematicians checked the result. Soon afterward, OpenAI presented ten more results involving open mathematical problems. Some were fully resolved, while others saw substantial progress. Insane.
A mathematical result can take a long time to verify and debate. But the fact itself is already difficult to ignore: models are entering areas where no ready-made answer exists.
Examples like these communicate the speed of what is happening better than another percentage increase on a chart. When a model begins working on problems that have remained open for decades, the conversation moves far beyond text and image generation.
The cost of experimentation
A model at roughly the level of GPT-5 can now be run without much difficulty on hardware costing $20,000–30,000. That is colossal progress. A year from now, today’s models may already be another order of magnitude cheaper.
The absolute price is not even the most important part. The direction matters more: capabilities that recently existed only inside the largest laboratories are gradually becoming available to small teams and individual people.
These models already cover an enormous range of applied tasks: generating images, video, and text; writing software; analysing data; finding solutions; and optimising processes. A single system can participate in almost the entire process of creating a product, from the initial idea to a working prototype.
In the past, money, specialists, and months of work stood between an idea and the ability to test it. Increasingly, the main constraint is the person’s ability to understand what they actually want to build and why.
A fragile hope
Whenever I think about this, it becomes difficult to imagine where we will eventually arrive. Of course, one can argue that all of this will lead us somewhere terrible — and it probably will. Yet what is happening also gives me some hope and optimism. Because if these breakthroughs disappear as well, we will be left with military conflicts and senseless political decisions that are reliably driving us toward misery.
I have a fragile hope that all of this might somehow shake up the current world order and transform society into something new. Perhaps it can give people more autonomy and more room to create.
Now, if you want to test an idea or start a project, you no longer need to recruit people and assemble a development team first. You can sit down and produce an imperfect but working prototype in a single evening. It becomes much easier to understand its usefulness and evaluate its prospects. That is an incredible luxury that did not exist before.
Of course, lowering the barrier works in both directions. Creation will become easier for everyone, so the simple fact that someone made a product will matter less. Taste, direction, the ability to notice real problems, and the discipline to finish what was started will become more valuable. The tools will be more accessible, while the competition will become harsher.
There is no grand conclusion here. I simply wanted to say that extraordinary changes are coming, and their scale is still too difficult to evaluate. We need to prepare for them. There will be no stability. And I cannot confidently say that this is a bad thing.