Should I route everything through OpenRouter?

The same model costs 6.8 times more on one route than another, and the router’s automatic choice is neither. Every route in that table serves the same weights. The cheapest cold price is 6.7 cents per million tokens and the dearest is 45; the first-party API sits in the middle. Automatic routing landed at 10 cents, better than most, and it did so by sending repeated prompts to different upstreams: over the study, the same request shape reached Baidu, CoreWeave, SiliconFlow, Parasail and a dozen others. That is the convenience working as designed. It also means the route your production traffic takes next week is not the one you benchmarked this week.

What a router hides is not the price, it is the three things next to it. First, precision: the routes in the table are labelled FP4, FP8 and BF16 because the hosts chose different quantisations of the same model, and a model name on a router is not a promise about which one you get. Second, caching: a repeated prompt only hits a cache if it lands on the same upstream with the prefix still warm, and the measured cache share ranged from 53% to 100% by route, which on an agent workload is the difference between the two lines on the previous page. Third, reliability: one route failed 55 of 100 calls in this sample, and a failed call that is retried is billed twice. None of these appear on the price page, and a router’s fee, a few percent on top of the upstream rate, is the smallest of the four.

A router is right when you are choosing; direct is right once you have chosen. While you are still deciding between models, or you need fallbacks, or you are shopping the open-weight hosts, one key that reaches everything is worth its fee and then some. Once one or two models carry most of your traffic, going direct buys you the things a router cannot sell: a cache you control the hit rate of, the batch tier at half price, committed-spend and provisioned-throughput pricing at volume, a single data-handling agreement rather than one per upstream, and a fixed precision. The pattern most teams settle into is both: direct for the workhorse, a router for the long tail and the experiments.

Where the real discounts are, in order of size. Open weights: a model in the same capability tier served from open weights is routinely twenty times cheaper than a frontier proprietary one on the list, before you ask whether the task needed the frontier at all. Caching: a tenth of the input price on every re-read, which on an agent workload is most of the tokens. The batch tier: half price for anything that can wait a day. Not paying for the frontier when the task does not need it: the newest flagship from each vendor lists at two to five times its own previous tier, and up to fifty times a small model from the same vendor. And time: list prices fall, and a rate negotiated last year is worth renegotiating. Which host actually wins on your prompts depends on their shape, the report shows the winner changing between 1K, 10K and 100K inputs, so the last discount is measuring with your own traffic before signing anything.