About eight months ago, I stopped thinking of Claude’s product problems as isolated bugs.
I concluded that Anthropic probably sees itself as a model company first. Claude.ai feels like a place to expose the model, rather than a product the company is determined to make excellent.
I use Claude often enough that I could write a separate note called “Ninety-Nine Things Broken in Claude.”
At one point, I could be speaking into the mobile app and the phone screen would go to sleep. I had to keep touching the screen while recording a long voice note.
How does that survive inside a product made by one of the leading AI companies in the world?
The voice transcription also struggles with my accent more than ChatGPT does. Search, context management, mobile behavior, desktop behavior, and other basic interactions repeatedly get between me and the model.
The strange part is that the models themselves can be excellent. Claude can reason through difficult problems, improve writing, and work through code. Sometimes I prefer its answer.
But using the model and using the product are two different experiences.
ChatGPT seems to take the product layer more seriously.
I can walk, speak for several minutes, and trust that it will understand most of what I say. I can listen to the response and continue the thought. I do not have to keep checking whether the screen has gone to sleep or reshape the way I speak so the transcription can understand me.
That matters because voice is how I produce a lot of my thinking. When the product works, I stop thinking about the product.
Anthropic may have good business reasons for this. Its real business may be APIs, enterprise customers, and selling access to the models. Claude.ai may be mainly a demonstration of what the models can do.
But models are still a means to an end.
Even AGI, if it arrives, will have to be served through something: an interface, an agent, a device, a voice, a workflow, or another kind of product.
Somebody still has to make that product work.
People keep saying software is solved because models can now write code. But software is more than producing code. It includes the interface, the search, the mobile experience, the reliability, and all the small decisions between the model and the person trying to use it.
The products these AI companies build for themselves may be the best evidence of whether software is actually solved.
Claude has shown me that an exceptional model can still sit inside a product that feels secondary.
And if models eventually become cheaper, more common, and harder to distinguish, the product may be the part that remains.
Part of the Software That Works series.