Comment by adamsmark

Local isn't viable yet on an economic basis, API costs are so low that you're better off taking advantage of the bonanza. As local models become more performant, so too will the ability of providers via Openrouter be able to offer them cheaper than your likely payoff period for a $4K Mac Studio 128GB. e.g Gemma 4 31B is impressive, but it costs practically nothing via Openrouter. Given that there are a ton of providers for open models, I doubt there's any subsidy going on because the providers are faceless and interchangeable.

At least, that's my theory.

The big advantages of local on a business level are:

- Freezing your model's exact settings once you've locked in some kind of workflow that works just fine. - Guarding against insane token usage from LLMs who have been told to never stop until they figure out the solution OR setting up an LLM run incorrectly. (The last one happened to me with Gemini 3.1 Pro) - PII or some need for on-premise only LLMs.