OpenRouter Made Your AI Calls Faster, and About 7% Pricier, Without Asking

On October 28, OpenRouter changed how it picks who answers a request. OpenRouter is a service that sells access to hundreds of AI models through one account, and most of them are run by several hosting companies, called providers. Until now, when a request didn't say which provider to use, OpenRouter mostly sent it to the cheapest one. The new default weighs price against speed, rate limits (a provider turning a request away because it's full), errors, and whether tool calls come back well formed.

OpenRouter's own test says the trade works, and that it costs a little. Over one week it split more than 24 million requests between the old and new router. With the new one, the first word of an answer arrived 16.9% sooner, output streamed 16.3% faster, and the share of requests that hit a rate limit fell from 7.1% to 3.8%. The price per million tokens rose from $0.301 to $0.324, or 7.3%. Per request, cost rose 6.0%. These are OpenRouter's numbers on its own traffic. No outside test has checked them yet.

The change is automatic. Any request that doesn't set a provider sort or a provider order now uses the new router, with the same API key and the same model names. A conversation stays with its first provider so the prompt cache, the saved copy of earlier text that makes repeat calls cheaper, keeps working.

Jon's agents call OpenRouter for checks like Jev, TypeSafe's yes-or-no judge model, and for image work, and the fleet has measured provider prices on it before. A check of the fleet's own scripts found none that set a sort or an order, so those calls moved to the new router on October 28 without anyone changing a line. For agents that make many calls in a row, faster first words add up. So does 7.3%.

How to use it

• To keep paying the lowest price: add "provider": {"sort": "price"} to the request, add :floor to the end of the model name, or set the default sort to price in account settings. Any of these skips the new router.
• To pick providers yourself: "provider": {"order": [...]} tries them in the order given. only and ignore still narrow the list, but the new router picks among what's left. The provider routing guide has every option.
• To compare costs fairly, compare whole conversations, not single requests, because each conversation stays pinned to one provider.
• The traffic routing page explains "effective price": what a provider actually charged over the last 30 minutes, after cache discounts, which can differ a lot from its list price.

What nobody knows yet

How much of the 7.3% shows up on any one customer's bill. The test covered a sample of open-source models over one week, before providers had time to adapt. And the quality check mostly tests the shape of tool calls today, so it can't yet tell a well-formed wrong answer from a right one.

What to watch

OpenRouter says the gains came before providers tuned for the new scoring. The next read is a month of real bills: whether the price gap shrinks as providers compete on speed instead of price alone.