OpenRouter early investors recount the investment process.

Author: Menlo Ventures

Compiled by: Jia Huan, ChainCatcher

Today, OpenRouter announced that it has reached an acquisition agreement with Stripe. This comes just three years after OpenRouter's official launch in 2023.

OpenRouter was initially launched as a "unified interface for LLM" and only supported four models at the time: GPT-3.5, GPT-4, Together's GPT NeoXT, and Cohere xlarge.

The company was founded on two core judgments: First, the scale of AI usage will eventually be enormous and permeate all fields; second, a large number of different models will emerge in the market, each with its own advantages and disadvantages, and users will choose different models according to their different needs.

As it turned out, both of these judgments far exceeded expectations at the time.

Since its launch, the number of tokens processed by the OpenRouter platform has increased approximately 30,000 times, currently exceeding 4500 trillion tokens annualized, and the scale of spending on the platform has reached an astonishing level. At the same time, the number of models supported by OpenRouter has increased from the initial 4 to over 500.

Chart: Growth in OpenRouter Token usage from its inception to its acquisition.

Menlo Ventures is fortunate to be a part of this journey. In March 2025, we participated in OpenRouter's seed round funding through the Anthology Fund, which we established in partnership with Anthropic.

Alex Atallah, founder and CEO of OpenRouter, previously founded OpenSea, which was once valued at $13.3 billion. His co-founders include Louis Vichy, a tech whiz he met on Discord, and Chris Clark, a highly effective COO.

In May 2025, we led OpenRouter's Series A funding round, with Matt joining the company's board of directors and Deedy serving as a board observer. Earlier this year, seeing OpenRouter's rapid growth in customer numbers and revenue, and the company's product roadmap building stronger "model intelligence" capabilities around model selection and evaluation, we continued to invest in the Series B funding round.

In the tech industry, an idea often takes years to go from being the judgment of a few to becoming an industry consensus.

Just a few weeks ago, this happened: from Ramp to Cursor, more than 10 companies launched their own model routing products almost simultaneously. In just a few years, OpenRouter has become one of the most important companies in the AI era.

Photo: Group photo taken when OpenRouter's Series A funding round was confirmed as the lead investor.

At first glance, Stripe may not seem like the most natural acquirer of OpenRouter, but the two companies are actually strikingly similar.

Both simplify complex transaction processes through a directly accessible API, charging a percentage of the transaction fee. However, OpenRouter handles AI models.

According to Stripe's consistent statement, the two companies, after merging, are still doing the same thing: increasing "Internet GDP".

In fact, more than a year ago, OpenRouter had already called itself "the Stripe of the LLM world".

The core value of OpenRouter

OpenRouter was one of the first companies Deedy came into contact with after joining Menlo in 2024. This company is almost exactly at the heart of our AI infrastructure investment logic.

In his " 2024 Enterprise AI Report ," Menlo proposed two assumptions that must be made for investing in OpenRouter: AI spending will increase significantly, and developers will not use just one model, but will use multiple models simultaneously.

Image: Menlo's initial email to OpenRouter

As people who also write code and use these models in practice, we realized early on that there are very significant differences between different models in terms of cost, latency, and performance.

For example, when you are doing a simple NLP task, such as recognizing entities from text, you don't necessarily need to call a cutting-edge, powerful model like Fable.

The problem is that if users need to go to the official website of each model company, register an account, create an API key, keep the key safe, adapt to the slightly different API interface specifications of each company, and finally manage all the models themselves, the whole process will be very cumbersome.

A unified model gateway may sound simple, but in reality, it's a much more complex infrastructure problem than it appears. Few people are truly willing to build and maintain such a system themselves in the long run.

The venture capital industry often discusses "moats," and usually thinks of technological barriers first. But OpenRouter possesses a very typical scale moat.

The more users OpenRouter has, the better it can predict model demand and handle larger loads; it is also easier to sign larger contracts with model labs, thereby obtaining a more stable supply and demand of tokens.

Ultimately, this creates a cycle: new model labs, in order to gain distribution channels, will also want to prioritize logging into OpenRouter.

We also observed another trend.

With the increasing popularity of Vibe Coding, the number of software startups has grown rapidly. For a product aiming to enter the enterprise market, those that have already won the approval of the company's internal developers are often the ones that are ultimately purchased by enterprises.

Anthropic, OpenAI, xAI, Cursor, Cognition, ElevenLabs, Lovable, and Fireworks all follow this pattern: they win over developers first, and then enter the enterprise market.

The same applies to OpenRouter.

Since our investment:

OpenRouter is currently processing 30,000 times the number of tokens it did at launch, and has maintained a monthly growth rate of about 33% over the past three years, doubling on average every 11 weeks.

The model market has also expanded rapidly. Excellent open-source models such as DeepSeek, GLM, and Kimi have emerged in China, along with models from companies like Grok, Meta, and Thinking Machines. Currently, OpenRouter has integrated over 500 models from more than 80 model providers, serving approximately 10 million users.

Many important new models will be prioritized for release on OpenRouter, including models from OpenAI, X, and Meta. Mark Zuckerberg, who rarely tweets or speaks out for other products, specifically announced the Muse Spark's release on OpenRouter, as did Elon Musk. OpenAI will also offer exclusive discounts on models like Terra and Luna through OpenRouter.

OpenRouter's product-driven growth model has also successfully translated into the enterprise market. Its enterprise product sales cycle is among the fastest we've ever seen. Enterprises can use this product to centrally allocate model resources, control access permissions, and manage internal AI budgets.

Because OpenRouter can negotiate contracts with different model providers in a unified manner, it can provide extremely high service availability even with the most cutting-edge models.

How did OpenRouter get to where it is today?

Anyone who has been running a startup for a long time will tell you that finding product-market fit, or PMF, is never a straight line.

The same applies to OpenRouter.

Its story actually begins on April 5, 2023. At that time, the team launched a Chrome extension called Window, which allowed users to call multiple models simultaneously in different chat applications on the Internet.

Their initial goal was to prevent users from being locked into a particular model vendor, while also eliminating the need to hand over their API keys to different applications to use different models.

The design of this product was inspired by crypto wallets, which is also related to Alex's previous experience founding OpenSea. At that time, Window supported a total of 4 models.

On April 24, 2023, the name "OpenRouter" first appeared in the Windows GitHub repository.

About a month later, they integrated the first batch of Anthropic v1 models, began automatically assigning requests to the appropriate models based on the prompts, created the now-famous model leaderboard, and started using the name OpenRouter.

On August 10, 2023, the team officially renamed the product OpenRouter, with the slogan "A unified interface for LLMs".

At that time, OpenRouter processed approximately 3 billion tokens per week. The leaderboard model back then was completely different from the model we are familiar with today.

The team then launched Playground, which allows users to receive responses from multiple models simultaneously through a single chat interface.

By November of that year, OpenRouter supported 52 models, connected to more than 2,000 applications, and processed approximately 8 billion tokens per week.

At this point, they finally found PMF.

The Future of Model Routing

Contrary to popular belief, OpenRouter's core product is not simply "helping users route tokens," but rather becoming the most user-friendly AI gateway.

OpenRouter does offer an Auto Router that can automatically select a model, but most developers use OpenRouter primarily to access different models through a single entry point and then decide how to route the models according to their own needs.

Recently, the rapid increase in corporate spending on LLMs has become a real problem for companies such as Uber, Coinbase, and Microsoft.

Therefore, the "model router" sounds very appealing as a cost-reduction solution.

Since an AI agent breaks down a task into many different sub-tasks, why use the most expensive model for each task? Simple tasks can be handled by lower-cost models.

In the past few weeks, the entire industry seems to have suddenly realized this simultaneously. From Ramp to Cursor, more than 10 companies have launched their own model routers.

However, the problem is that selecting a model solely based on the Prompt itself is not a particularly effective method.

In an agent-based scenario, a task can take a very long time to complete. Determining which model should handle a particular request requires understanding a significant amount of context.

For example, a simple command: "Find this file in the code repository."

It might only require a cheap, simple LLM, or it might have to call a state-of-the-art model. It depends on the size of the codebase and the context that has been built up by previous tasks.

In multi-step agent tasks, if the router selects the wrong model at any step, the cost of this error will continue to amplify in subsequent steps, eventually leading to a significant decline in the overall quality of the task results.

Therefore, the foundation of a truly differentiated model routing product is not a simple "model selection algorithm," but an excellent unified API and a sufficiently large real user base.

It is these users who have enabled OpenRouter to gradually accumulate an extremely large dataset, almost unnoticed by the outside world. This dataset includes what prompts users submitted, what models these prompts ultimately used, what contexts were used during task execution, and what results were achieved.

This is where model routing truly matters.

In a real production environment, OpenRouter can help businesses control costs by maintaining performance as much as possible through more reasonable model routing.

In the future, developers will not need to build complex evaluation systems themselves, nor will they need to continuously modify the Prompt for different models.

You might simply need to log into the backend and see a message like this: "A portion of your codebase is primarily used for summarizing. Switching from GPT 5.6 Sol to Muse Spark could save you $100,000 annually. We've automatically performed the relevant evaluations for you."

This is the true future of model routing.

From payment infrastructure to AI infrastructure

Stripe and OpenRouter have very similar development trajectories. Both companies first won over developers before gradually entering the large enterprise market. Their product design languages are also very simple and direct.

Andrej Karpathy once called OpenRouter a "switch" for AI. Stripe does something similar: it has become that "switch" in the payment processing system.

This acquisition is one of the earliest major deals in the AI era's infrastructure sector, and it won't be the last. The infrastructure for managing models, costs, and computing power is gradually taking shape. This infrastructure is enabling the formation of a new generation of giant companies at a faster pace than ever before.