It's a few hundred bucks per month for now, but that's not going to last. At some point, the industry is going to pivot towards tracking token-based productivity because it's not going to be cheap forever unless FOSS models catch up.

Please don't call open weight models FOSS models - that's actually very wrong, unless you actually have all the training data and can modify the data and training methodology to retrain the model yourself.

FOSS models have effectively caught up wrt. scale, see e.g. the latest DeepSeek V4 series - but they still require major hardware resources (hundreds of gigabytes of RAM for a very lean deployment targeting single- or few-users inference) to run at acceptable throughput.