I've seen three or four others building routers. How hard is that really? The folk I spoke to needed only a few millions to build it (they claim). I've seen prototype of some too that looked pretty simple. Maybe the whole thing is driven by enterprise wanting the external vendor support?

The software is "easy", it's the business and operations that are hard: scaling the infrastructure reliably; finding revenue sources with secure margins; deflecting regulatory and risks responsibility; billing and collecting on massive low-margin high-volume transactions; enterprise sales; etc

It's a business line more suited to finance, law, sales, and accounting people than tech people and a pretty laborious one. That's often the case when the tech looks "easy" but the sector only seems to have a few big winners.

My surface-level assessment is that building a router is easy, but routing efficiently is difficult. For the latter problem, traffic provides signals which you can use to improve routing rules, creating a feedback loop. More traffic => more feedback => better routing. So a lot of the value of Openrouter would be that they are the most adopted router (I think?) and so likely have the most data to (potentially) produce the best router. I'm only speculating.

It's hard.

- They support 400 different models, 80 different providers, and an unlimited number of new custom agents. Every single model, provider, and agent, has its own weirdness that has to be accounted for. Tool calls change by model. Effort changes by model. Backend APIs (messages, responses, etc) change by provider. There are thousands of specific tweaks, fixes, hacks, that need to be implemented to make this thing "just work". And you have to keep updating it all, weekly.

- On top of that, they support providers running in multiple countries, which increases the legal, logistical, financial, and networking complexity.

- On top of that, they need to figure out pricing and deals with providers, and ensure the providers allocate the necessary networking and compute capacity. They need to work with providers directly to troubleshoot and fix all kinds of issues, from networking to application.

- On top of that, they implement multiple kinds of request routers. An auto router to route your request to the best general model, a fusion model to attempt the request on multiple models and pick the best response, a pareto router to route requests to the best coding model for your request.

- On top of that, they build custom features that businesses and users want, like Data Loss Prevention (which I'm blown away they actually provide for free). Lots of very useful business functionality for managing not only what model and provider to use, but also limits on usage, filters, etc. They also implement SSO, prompt injection guardrails, logging/auditing, workspaces, etc.

- On top of that, you have to be very good at just implementing HTTP APIs. Most people aren't good at it. API design is hard, HTTP is way more complicated than it seems, network traffic shaping is a black art.

- On top of that, running applications and infrastructure, and scaling it 10x every year, is a subtle yet critical skill. No matter how good your code is, it's pointless without working hardware/running apps.

- On top of that, they provide a generous free tier, which has to be subsidized, and is probably only partially subsidized by providers.

Why pay for this rather than build one? Because there's no sense in building one. There's a reason restaurants don't build their own factories to manufacture their own pots and pans. Their business isn't selling pots and pans, it's selling people cooked steak. It's a bad idea to waste time, effort, and money trying to make things you aren't going to sell.

It's a fairly high scale, but yeah I don't think it's terribly hard. Negotiating rates with the providers seems like it'd be the hardest part to me.

Probably they want to use the userbase, not the technology? There was a heavy emphasis - too heavy for my taste so it bounces me back - about that they will do what is the best for the user. I think that was already there, Stripe was not needed for that. This is more for the benefit of others.