The work that stops
There will be more subscription-based programmers. Their ability to do the job depends solely on a platform: OpenAI, Anthropic Claude, and the others that follow. The key, the model, and the meter are the job. Take the subscription away, and the work stops.
The login is the skill. The model writes the function, names the file, and explains the diff. The person approves it. When the key expires, the rate limit hits, or the model is retired, that person is back at an empty editor.
What the subscription sells
The subscription sells a translation. A sentence goes in. A file comes out. The file compiles. The demo runs.
That translation is the part that got cheap. It is also the part that was never the job. A fluent file can satisfy the sentence and miss the person.
The problem and the expectation
Capturing is writing down the problem in the form a person could not quite say. Understanding is knowing which sentence was the product. Analyzing is finding where the expectation breaks: a charge that happens twice, a session that belongs to the wrong person, an account that cannot be left. Managing is keeping that expectation still when the draft changes and a second person arrives.
None of that is in the translation. The platform can draft the file. It cannot tell which expectation was the product.
The meter is not the problem
Price, context, and policy belong to the platform. A programmer who works only through the subscription inherits all three. The problem did not get smaller because the window did. The expectation did not get clearer because the model was confident.
Two people can pay for the same model and ship two products. The difference is who understood the problem.
The test
Take the subscription away. Can this person still say what the user was trying to do, what the product agreed to do back, and where that agreement will break?
If the answer lives on the platform, the programmer was renting the job.