OpenRouter used to send your request to the cheapest provider that could take it. Its new default charges a little more for a better chance of a fast, clean answer.
OpenRouter Rewrote the Rule That Picks a Provider
On October 28, OpenRouter described the largest change to its provider routing since the company started. The default no longer sorts providers by input token price. It ranks them by something the company calls Routing Cost.
That sounds like accounting trivia. It isn't. About 77% of requests on the platform go to models that several providers serve, so the default rule decides where most traffic lands, which provider gets paid, and how long you sit waiting for the first token.
If you call a model through OpenRouter, you've been relying on that rule whether you knew it or not.
What the Old Default Got Wrong at 77% of Traffic
Under the old system, the lowest input price won. A provider with worse hardware could take the traffic as long as its headline rate was a cent cheaper than everyone else's.
Rankings were static. Nobody saw a provider degrade in real time, so a bad afternoon at one datacenter could follow you around for hours.
The incentives did the rest. Providers competed by cutting prices, which squeezed the money they had for infrastructure. Some played a pricing game: advertise a low input rate and recover the margin on cached output. Cheap on the way in, expensive on the way out.
Every one of those outcomes landed on the same person. You.
Routing Cost, Explained Without the Math
The new model tracks actual price, latency, throughput, and error rate for every model on every endpoint, then estimates the cost of getting one good answer.
Two pieces feed that estimate. The first is the price of a single attempt. OpenRouter uses the effective price after adjusting for recent cache hit rates and the input-to-output ratio, which closes the gap that let providers game list prices. Waiting time gets folded in too, with throughput and time to first token converted into money at rates set per model.
The second piece is the chance the attempt works. Reliability counts, and provider-side errors get penalized hard. A provider whose error rate runs far above its peers loses traffic until it recovers. Quality counts as well. For now that mostly means whether tool calls come back in the right format, with plans to widen it later. Providers with clearly low benchmark accuracy get demoted to a small slice of test traffic.
The arithmetic is easy to follow. Fail one attempt in five and you pay 1.25 times the sticker price for every good answer you get.
Reliability Now Moves Traffic, Not Just Price
A router that always picks the single best endpoint would be fragile. One bad hour and every request suffers. OpenRouter doesn't work that way. What happens when a provider has a bad week?
Endpoints inside a competitive band share most of the traffic. Fall slightly behind and you get less of it. Drop out of the band and volume falls quickly, sometimes inside an hour. Each provider keeps a small base of traffic for observation, so the system notices when someone improves.
That last detail matters more than it looks. Without a floor, a provider that stumbles once would never get the data it needs to prove it fixed the problem.
The 16.9% Latency Win and the 6% Bill Increase
OpenRouter measured the results itself, and the numbers are its own. Time to first token fell 16.9%. Throughput rose 16.3%. Capacity errors, the HTTP 429 responses that appear when a provider has no room, dropped 67%.
Cost rose 6.0%.
One caveat comes from the company: these figures predate any provider tuning for the new rules. Providers that adjust for Routing Cost should move the numbers further.
For a team running a chat product, 16.9% off first-token latency is the change users feel in their hands. For a team watching a bill, 6% is the change finance feels. Both are real, and they point in opposite directions.
Caching Still Decides Where You Land
Routing remains cache aware. While a conversation's cache is warm, the session stays on the same provider, since moving breaks the cache and raises the price of every following turn.
That's a sensible constraint and also a limit on how much the new algorithm can do. A session that starts on a mediocre endpoint tends to stay there until the cache goes cold. In plain terms, the router stops shopping for a better price once your conversation has warmed up its context, because re-reading that context elsewhere costs more than the move saves.
Teams that want the old behavior have options. Pinning a provider and sorting by price restore pure price routing for your account.
How to Test the New OpenRouter Default
Try the default for a week before touching the controls. Compare your bill against last month and watch two things: how often you see a 429, and how quickly the first token arrives.
If your workload is batch jobs where a retry costs nothing, the old price-first logic may still be the better deal. If it's anything a person waits on, the trade looks different.
The bigger question is whether other routers follow. OpenRouter's argument is that paying a little more for reliability beats paying the lowest sticker price and eating the retries. The 429 drop is the part you can measure on your own traffic, and it's the number worth checking first.






