Open-Weight vs Proprietary
Whether a model’s weights are published (so any host can serve it) or kept private and sold only by the vendor.
The open-weight vs proprietary distinction is about who can serve a model. Proprietary models — Claude, GPT, Gemini — are sold only by the vendor that trained them, so there is one authoritative price per model. Open-weight models — Llama, DeepSeek, Qwen — have published weights, so many hosts can run them and each sets its own price, and you can even self-host.
For pricing, this matters in two ways. First, an open-weight model has no single price: the figure you see depends on which host you pick. For the Llama 4 models, for example, this explorer lists a third-party host’s serverless rate (Together AI, which has since stopped serving them) for reference and labels the host explicitly — a different host may charge more or less. Second, open-weight models are often dramatically cheaper per token because hosts compete on price, though you trade away the vendor’s managed tuning, safety, and support.
This explorer labels third-party-hosted prices with a host badge so you always know whose rate you are looking at, and never presents one host’s price as the model’s only price. When evaluating an open-weight model, compare hosts, check throughput and reliability alongside the token rate, and factor in the engineering cost of self-hosting if you go that route. The cheapest sticker is not always the cheapest total.
FAQ
Why does an open-weight model have different prices?
Because its weights are public, many hosts can serve it and each sets its own rate. The price you see depends on the host — this explorer labels which one.
Are open-weight models always cheaper?
Often cheaper per token, since hosts compete on price, but you trade the vendor’s managed tuning and support. Weigh throughput, reliability, and self-hosting effort alongside the rate.