One cost three cents.
Most of my building runs through workers now. I say what I want, Nav (my Claude setup) turns it into a spec, and a worker on my desktop builds it while I get on with something else.
The real question was which model gets to be the worker.
For a while it was MiniMax M3. Cheap monthly plan, did the job. But it came with a five-hour window. Hit the limit, wait. Not great when the whole point is that work keeps moving while I'm not watching.
So in September I tried DeepSeek. Same task file, word for word, sent to both.
Both came back with tests passing. Then I compared the two changes against each other, and they were identical, apart from one line of notes. 82 out of 82 tests green on both.
The DeepSeek run cost three US cents at peak. About one and a half off-peak.
I expected close enough. I did not expect identical.
One thing worth knowing if you try this - Claude Code shows a cost for every run, and for DeepSeek that number is made up. It doesn't recognise the model, so it guesses, and on one run the guess was 39 times too high. The token counts are fine. Work the cost out from those.
DeepSeek has been my default worker since 21 September. I cancelled the M3 plan a few days later.
I think the lesson isn't really "DeepSeek is better". Once the spec is good enough, the worker matters a lot less. The spec is the product, the model is just who shows up to build it.
Now I just have to write specs that good every time. That part is still on me.