OpenAI just released their first open source models in 5 years. I've been testing them and honestly, this is bigger than people realize. GPT-OSS comes in two sizes: 12B and 20B parameters. The smaller one runs on a laptop with 16GB RAM. The bigger one needs an 80GB GPU but still fits on a single card. Here's the thing that caught my attention. These aren't just smaller versions of existing models. They can reason through problems and call tools while thinking. No agent frameworks needed. I ran some benchmarks and these models are punching way above their weight. They're competitive with GPT-4 and O1 on many tasks despite being much smaller. 90% on MMLU is no joke. But what this really means is the economics just shifted. Hosting providers are already racing to the bottom on pricing. I'm seeing 5 cents per million input tokens. Compare that to proprietary models at 10-20x the cost.