Every lab sells its models through its own counter: account, payment method, key, console. If you want to compare three models from three different companies on your own cases, the paperwork costs more than the test. OpenRouter exists to remove that friction: one account and a single key, with 400+ models from Anthropic, OpenAI, Google, Meta and the open-weight labs behind it, all with the same call format and the same balance.

It is the tool we use ourselves: this site’s editorial pipeline calls its models through OpenRouter, and the comparisons in our “we tried it” series run there.

How billing works

You buy prepaid credits and each request consumes them. The price per million tokens is the provider’s list price, with no markup: the fee sits upstream, 5.5% when you purchase credits (checked on 19/07/2026 on the pricing page). There is no subscription: if you don’t use it you don’t spend, and the credits stay put.

Two details worth knowing before you reach for your wallet. Many open models have a free variant (the :free suffix in the name) capped at 50 requests a day, rising to 1,000 after a one-time $10 credit purchase. And every API key can carry a spending cap with a daily, weekly or monthly reset: past the limit, requests fail instead of quietly spending on, which is the right guardrail if a script gets away from you.

The playground, for people who don’t code

The least-told part of the value: at openrouter.ai/chat you can compare up to five models side by side on the same question, from the browser, without writing a line of code. The method this site keeps recommending, your own cases over leaderboards, happens exactly there: paste in a real case, look at five answers next to each other, decide with your own eyes.

The limits, said plainly

Your request makes one extra hop: it goes through OpenRouter’s infrastructure and then through whichever provider the routing picks for that model, so there is some added latency, and the same model can be served by different providers with slightly different behaviour. On the data side, OpenRouter logs request metadata by default but not your prompts, and in the settings you can exclude providers that train on your data or retain inputs from the routing; it is still one more actor in the chain, and for genuinely sensitive material the usual rule applies: decide first what never gets pasted.

For a production project with high volume and contractual requirements, a direct relationship with the lab (Anthropic, OpenAI or Google APIs) makes sense again: contracts, data guarantees and one less intermediary. OpenRouter wins where breadth matters: trying, comparing, switching model with one line.

The prices quoted here carry their verification date in the table above; the live price list is at openrouter.ai/models.