GPT-5 has dropped, to much rejoicing, from many people.

I will go slightly against the grain and say this is a disappointing release. Rather than a substantial new architecture innovation or data improvement, they just tacked a router in front of their models. Now, every prompt will be routed to the model that OpenAI judges is best suited to respond. LLM routers are not novel in any way. OpenRouter–my API provider of choice for a while–has had a model router for a while now. It’s a useful tool to be able to automatically send prompts to the model best suited to the task, but it’s not a novel idea or a huge improvement.

Instead GPT-5 is more of a UI improvement. For a while now, Altman has said publicly that OpenAI wants to do away with the model picker. This makes total sense. Most people don’t have the time or desire to research all the models to really grok their capabilities, read the benchmarks, and know what model is best suited for their task. Most people want something that Just Works™. In this sense, GPT-5 is a huge upgrade to the ChatGPT.com UI. Most people will see huge improvements because most people were probably just using regular old GPT-4o and were disappointed when it couldn’t accurately perform any remotely complex task.

But for substantial model-level improvements, there’s not much here. Certainly not enough to justify the massive amounts of hype. Again, OpenRouter has had their own model router for years at this point. If you’re using a single interface for every task and you aren’t very good at manually picking your own model, then you’ll probably see an improvement. If you already use different models in different interfaces (one model for coding, another for chatting, etc.), or know the capabilities of each model and switch between them, then this isn’t much of an improvement.