Open-Weight Models Reduce Inference Costs to One-Fifth of Closed Models

Open-weight models standardized by intelligence cost roughly 20% of comparable closed models. Self-hosting on rented hardware can lower costs to $0.12–$0.35 per million output tokens at full utilization, making older GPUs like the A100 more economically efficient than H100s for specific sparse architectures like gpt-oss-120b.