Stripe expands its playground. After optimizing financial flows, the payments specialist is now tackling the consumption of AI.
With the acquisition of OpenRouter, it aims to act at the heart of the machine to route requests, optimize tokens, and above all bring down the cost of models.
OpenRouter’s business is to abstract away the proliferation of AI models and providers. Instead of integrating the APIs of OpenAI, Anthropic, Google, Meta or other suppliers separately, a developer can pass through a single interface.
The platform can then determine which model to route a request to based on several parameters: price, speed, availability, reliability, or task complexity. This routing capability becomes especially important as companies increasingly combine multiple models within a single application.
OpenRouter lets developers access, through this single interface, more than 400 AI models from over 80 providers.
Although the amount of the deal has not been officially disclosed, sources close to the matter estimate it at a little over $8 billion, according to Reuters. In other words, the largest acquisition Stripe has made to date.
With OpenRouter, Stripe wants to control the AI bill
OpenRouter indicates that its infrastructure now processes more than 10,000 billion tokens per day and serves more than 10 million developers and companies. Its platform provides access to more than 400 models from over 80 providers. Nvidia, Zoom and Lovable are among its clients.
The rapprochement between the two companies is less surprising than it might seem. Stripe already optimizes its clients’ financial transactions by arbitrating between payment methods, authorization rates, or fraud-fighting mechanisms. With OpenRouter, it can apply a similar logic to AI consumption.
Which model to use? At what price? With what level of performance? And how much will this request really cost?
Questions that have become central as companies move from experimentation to large-scale deployments.
Stripe had already begun to position itself in this field before the acquisition, notably with Token Billing, a service designed to measure and bill the use of AI models. The acquisition of OpenRouter now provides the component that lets it act directly on this consumption.
The token, a new economic unit
For Stripe, the token is therefore set to become a new economic unit that must be optimized.
This approach takes on a particular dimension with the arrival of AI agents. A traditional application generates a relatively predictable number of calls. An agent capable of planning a task, using tools, consulting data, and calling on several models can, however, multiply the calls to LLMs.
The problem is then no longer merely choosing the best model, but selecting the model that is sufficiently capable at the best cost for each step of the process.
That is where OpenRouter’s positioning becomes especially interesting for Stripe.
For example, a company could use a high-end model for a complex task, a cheaper model for simple classification or summarization, and automatically switch to another provider in case of saturation or unavailability.
The router thus becomes a kind of AI expenditure management layer.
It’s not far from what Stripe has been doing for a long time in payments. A company does not necessarily worry about the complexity of the banking network behind its payment; it trusts Stripe to orchestrate it.
A market that goes far beyond OpenAI
The appeal of this position also lies in the fact that Stripe does not bet on a single model.
OpenAI, Anthropic, Google, Meta and the numerous open-source model providers are locked in a continuous battle over performance, prices, and capabilities. In this environment, betting on the layer that sits between models and applications may be more durable than backing a single supplier.
That is precisely the bet OpenRouter has made since its founding in 2023. Rather than trying to build a competing model, it positioned itself as an infrastructure that enables the exploitation of the abundance of models.
Its acquisition by Stripe can be seen as something of a validation of this approach.
