OpenRouter: our review

OpenRouter doesn't build or host any model itself. The company builds a routing layer: a single API key gives access to more than 400 models published by dozens of publishers — OpenAI, Anthropic, Google, Mistral, and a long list of open models — spread across more than 70 inference providers. It's a product for technical teams, designed to avoid managing a separate account and bill per model publisher.

A router, not a host

In practice, a request sent to OpenRouter is routed to the chosen provider according to a mode set by the client: balanced between price and speed, priority to speed, or priority to tool-calling accuracy. If a provider goes down or raises its prices, OpenRouter automatically switches to another one serving the same model — a safety net a direct contract with a single publisher doesn't offer. The price shown for each model generally matches the original provider's rate: GPT-4o mini, for example, stays at $0.15 per million input tokens and $0.60 for output, identical to the direct rate at its publisher.

The price of convenience: a commission, not a fixed rate

This convenience has a cost specific to OpenRouter: a 5.5% commission on the amount of each request on the standard offering, on top of the provider's rate. The free tier, meanwhile, gives real access to more than 25 models with 50 requests a day, enough to prototype an AI agent before settling on a final provider. The Enterprise offering negotiates the commission down for a committed volume, on quote.

Confidentiality, provider by provider

This is a question specific to a router, and it deserves to be asked differently than for a single host: OpenRouter states it doesn't itself use prompts or responses to train a model, but every request is passed to the chosen provider, which applies its own policy. The documentation provides two concrete levers: the data_collection parameter (allow/deny) excludes providers that allow themselves to train on the data, and the zdr option restricts the request to zero-retention providers only. Without these settings active, protection depends on whichever provider is picked by default — not on OpenRouter.

Where it falls short

Two further reservations. First, fully European data processing only exists on the Enterprise offering; by default, the privacy policy provides for a transfer to OpenRouter's US servers, outside the European Economic Area. Second, the 5.5% commission adds up on a large volume to the point of eventually outweighing the savings made by avoiding several separate contracts.

Our recommendation

OpenRouter wins over a technical team that wants to compare models or protect its production against a provider outage, provided it sets data_collection and zdr on every sensitive call. For very high-volume use or a strict requirement for European processing, a direct provider such as Scaleway or Hugging Face stays more predictable. See also the other model platforms in this comparison.