I was testing the AI through VSCode Copilot. I had a piece of code that didn’t seem efficient, so I tried using the latest (and most expensive) model available at the time: Claude Opus 4.6.
The result wasn’t bad; it gave me two suggestions, one more convincing than the other. But what surprised me was that the suggested code included a using statement at the top: using godot. Godot? The project is a Unity game. I imagine a godot library exists within a Godot context, but why assume that engine? Don’t newer models read the context or at least the file structure anymore?
I decided to try Claude Opus 4.5, same prompt. The result: only one solution, the most convincing one, without errors or unnecessary using statements.
I noticed something similar a few days ago with one of the latest versions of ChatGPT: 5-mini. The response wasn’t bad. However, it misspelled an important variable as JONLIN_PROFILE_DIR, when the correct variable was JOPLIN_PROFILE_DIR. Nothing serious, since the error was in the response description and not in the code itself. But quite strange, considering it’s one of the latest versions of ChatGPT.
My personal conclusion: AI seems to have reached its first ceiling. If Claude releases Opus 4.7, or OpenAI releases ChatGPT 5.3, it will be much later than expected and with very little improvement compared to their predecessors.