Short answer: Jordan Metzner says the models cost essentially nothing beyond the servers that run them; the 20b runs on most computers, and the 120b is close to GPT-4.1 or o3 on much cheaper infrastructure.
Watch this answer from Episode 7 (starts at 15:48) · Episode 7 notes and transcript
From Episode 7: How Daily Check-Ins Fix Remote Work, starting at 15:48.
Sam Nadler [15:48]: Let me queue you up. I know I think it was yesterday, OpenAI released two open source models. I know that was exciting for you. What does that mean?
Jordan Metzner [16:00]: Today, as we're recording this, it seems like OpenAI is gonna announce even more new models. We'll talk about that in just a sec. But OpenAI released what they call OSS one twenty b and OSS 20 b. And what this means is these are models that have essentially zero cost to run. So you're only gonna have to pay for the servers to run them versus the previous models from OpenAI were available only through OpenAI's APIs. And so with this open source opportunity, you can actually host the models on your own computers. And for the 20 b model, you can run that on most computers. And what that means is that you can, use things like have it program all night long while you're sleeping and just use the power of your computer. Or for the one twenty b model, it's a pretty powerful model. It's equivalent to almost 4.1 or even close to o three style model, but you can get it running on a much cheaper infrastructure, something like Groq, or I saw, AWS announced it.
Jordan Metzner [17:04]: And so what you're able to do is do things that have almost, negligible cost except running the server to run those performances. And so you basically get, the ability to do almost anything you want with ChatGPT at a cost of zero. And if that means, the cost is zero, it means you can scale up out, you how often you could run, these servers, and you can run them frequently and concurrently, and, you can go horizontal and vertical, but, you could just have a server read every single line of code nonstop, like, in your entire code base all night long. And, once it's done, you could have it, start again type of thing and just look for bugs. Right?
