The switchboard nobody sees - one endpoint sitting between eight million developers and every large language model worth calling.
FILED FROM NEW YORK: A 51-person company that builds zero AI models and somehow touches almost all of them. The plumbing under the boom.
Somewhere a chatbot needs an answer. The developer who built it never picked a provider for this particular call. They wrote one line of code pointing at OpenRouter, and walked away. In the milliseconds that follow, OpenRouter checks who is cheapest, who is fastest, who is actually awake, and hands the prompt to the winner. The app gets its answer. Nobody thinks about it again. That is the entire point.
This happens roughly a hundred trillion times a month. OpenRouter is not a model. It is not a chatbot you talk to. It is the quiet layer underneath - the one that turned sixty competing AI providers into a single drop-down menu, and then made the menu disappear.
By 2023 the model count was exploding. OpenAI, Anthropic, Google, Mistral, and a long tail of specialists each shipped their own API, their own auth, their own billing, their own quirks. Want to compare two of them? Write two integrations. Want a backup when one rate-limits you at 2am? Write a third. The AI gold rush had arrived with no shared road.
The tax on all that choice was paid in engineering hours, and it compounded with every new model release.
Lock-in was the business model, even when nobody said it out loud. Switching costs kept customers parked on one vendor. The friction was a feature - just not for the people building things.
Alex Atallah had already built one marketplace. As co-founder and CTO of OpenSea, he watched a single interface absorb a chaotic, multi-vendor ecosystem. With co-founder Louis Vichy, he made a similar wager on AI: the winner would not be the company with the best model, but the company that made every model interchangeable.
It is a deeply un-glamorous bet. There is no model to brag about, no benchmark to top, no demo that makes a crowd gasp. There is only a promise that nobody will have to think about you. In an industry addicted to spectacle, OpenRouter chose to be infrastructure.
Investors noticed anyway. a16z led a seed. Menlo led a Series A. Then Alphabet's own growth fund, CapitalG, led a $113M Series B - a fascinating vote of confidence from the parent company of one of the very providers OpenRouter routes around.
The interface is deliberately familiar - fully OpenAI-compatible, so most apps switch by changing a single base URL. What changes is what happens after the request leaves. OpenRouter shops it across providers, applies your rules on price and latency, and quietly fails over to a backup when something breaks. You keep the provider's pricing; you gain everyone's uptime.
One OpenAI-compatible endpoint reaching 400+ models from 60+ providers, with shared billing, auth, and usage tracking.
Each request routed by cost, speed, and availability - with automatic failover when a provider is down or rate-limited.
Discover and compare models using real-world usage data, then pick the best price-to-performance for the task.
Provider filtering, zero-data-retention options, prompt caching, and request-level data policies for enterprise needs.
WHAT YOU CAN DO WITH IT: swap GPT for Claude for Gemini without a rewrite, A/B-test models on live traffic, set a cheap default with a premium fallback, or just sleep through a provider outage you never knew happened.
Run-rate inference spend went from about $10M in October 2024 to north of $100M by May 2025 - roughly a 10x climb in seven months. Then the token counter took over the story. Weekly volume hit around 25 trillion at the Series B; monthly volume sits near 100 trillion. More than a million developers have called the API, and the user base now tops eight million.
The Series B cap table reads like a list of the AI stack itself: CapitalG, NVIDIA's NVentures, Snowflake, Databricks, MongoDB, and ServiceNow Ventures all wrote checks. When the data and infrastructure giants invest in the layer that sits above the model providers, it says something about where they think the leverage is.
Strip away the funding headlines and the mission is small and stubborn: no AI application should be married to a single model or a single company. Switching should cost almost nothing. Choice should be the default, not a project. OpenRouter is betting that a neutral layer - one that stays useful precisely because it never picks a side - becomes more valuable as the model landscape grows louder and more crowded.
Every week brings a new model, a new price cut, a new specialist that is briefly the best at one thing. For builders, that is either a gift or a tax - depending entirely on whether they have to integrate each one by hand. OpenRouter's bet is that the model layer keeps fragmenting, and that fragmentation makes a neutral router more essential, not less.
So go back to that request leaving the app. A year ago, the developer would have hard-coded a provider, crossed their fingers at 2am, and rewritten code every time a cheaper model appeared. Now the request just leaves. OpenRouter checks who is cheapest, fastest, and awake - and hands it off. The app gets its answer. Nobody thinks about it again. The plumbing held. That was always the whole job.
Official channels, demos, and where to read more