I had early access to GPT-5.5, and my conclusion is simple: this is a significant step forward.
It matters for three reasons.
First, it shows that the rapid progress of AI is far from over. Despite years of predictions that AI improvement would slow down, frontier models continue to become more capable, more reliable, and more useful.
Second, GPT-5.5 is simply a very strong model. It can handle increasingly complex tasks that previously required teams of specialists, combining reasoning, coding, research, design, and creative work into a single workflow.
But third — and perhaps most importantly — GPT-5.5 also reminds us of something we have learned repeatedly throughout the AI era: the frontier of AI capability remains uneven.
AI is becoming extraordinarily powerful, but it is not uniformly intelligent. Some tasks that seemed impossible only a year ago are now routine, while other seemingly simple tasks still expose surprising weaknesses.
The frontier is moving quickly. But it is still jagged.
A Simple Test of How Far AI Has Come
As AI models improve, it has become increasingly difficult to demonstrate each new generation’s progress with simple examples.
Many classic AI challenges — solving complicated equations, understanding text, writing code, or even performing tasks like counting characters in words — have become almost trivial for modern models.
So instead of testing isolated skills, I wanted to see how different generations of AI approached a genuinely complex creative and technical challenge.
Coding remains one of the strongest areas for AI, so I gave several models the same challenge:
“Build a procedurally generated 3D simulation showing the evolution of a harbor town from 3000 BCE to 3000 AD. The simulation should look beautiful and allow users to interact with and control the experience.”
I tested models ranging from OpenAI’s first reasoning model, o3 (released just over a year ago), to current open-weight models such as Kimi K2.6, and finally GPT-5.5 Pro.
I published every result in an interactive gallery so people could explore the differences themselves. Interestingly, I even used GPT-5.5 Codex to build the gallery page that displayed the experiments.
The differences were immediately obvious.
Most models could create something visually interesting. However, GPT-5.5 Pro was the only one that truly modeled the idea of an evolving civilization.
Earlier models tended to replace buildings or generate new visual elements over time. GPT-5.5 Pro created something closer to an actual simulation — a town changing, developing, and transforming across thousands of years.
Speed improved as well.
GPT-5.4 Pro required approximately 33 minutes to complete the task, while GPT-5.5 Pro completed it in around 20 minutes while producing a more sophisticated result.
The improvement was not just intelligence.
It was the ability to execute.
Models, Apps, and AI Harnesses
I have increasingly encouraged people to stop thinking about AI as a single product.
AI is not just a chatbot.
It is a combination of three interconnected layers:
- Models
- Applications
- Harnesses
Understanding the difference between them is becoming essential.
1. Models: The Intelligence Layer
The first layer is the model itself.
Examples include:
- GPT-5.5 from OpenAI
- Claude Opus from Anthropic
- Gemini from Google
These models provide the underlying intelligence — reasoning, language understanding, coding ability, image generation, and multimodal capabilities.
GPT-5.5 represents another major improvement in this layer, particularly with its ability to handle complex, multi-step tasks.
The most capable version, GPT-5.5 Pro, demonstrates how far frontier models have progressed.
However, raw intelligence alone is not enough.
A powerful model sitting inside a basic chat window is like putting a supercomputer behind a calculator interface.
The real transformation comes from combining models with better tools and workflows.
2. Apps: Turning AI Into Something Useful
The second layer is the application.
Most people experience AI through websites:
- ChatGPT
- Claude
- Gemini
These interfaces introduced millions of people to AI, but they represent only the beginning.
The next generation of AI applications is moving beyond simple conversations.
Desktop AI environments such as:
- Claude Code
- Claude Cowork
- OpenAI Codex
are becoming far more powerful because they allow AI to work directly with files, software, and professional workflows.
The difference is significant.
A chatbot answers questions.
An AI application helps complete tasks.
3. Harnesses: Giving AI the Ability to Act
The third layer is the harness — the tools and systems that connect AI models to the real world.
Harnesses allow AI to:
- write and execute code;
- browse information;
- analyze documents;
- create images;
- control software;
- interact with computers.
This is where AI begins to shift from a conversational tool into an actual digital worker.
The biggest advances in AI are no longer only about making models smarter.
They are about making AI capable of doing more.
