OpenRouter Early Investors Review the Investment Journey

Author: Menlo Ventures

Compiled by: Jiahua, ChainCatcher

Today, OpenRouter announced that it has reached an acquisition agreement with Stripe. It has been just three years since OpenRouter officially launched in 2023.

OpenRouter initially positioned itself as "a unified interface for LLMs," supporting only four models at the time: GPT-3.5, GPT-4, Together's GPT NeoXT, and Cohere xlarge.

The company was founded based on two core judgments: first, the scale of AI usage will ultimately be enormous and will permeate various fields; second, a large number of different models will emerge in the market, each with its trade-offs, and users will choose different models based on their varying needs.

It has proven that both judgments far exceeded expectations at the time.

Since its launch, the number of tokens processed by the OpenRouter platform has grown by approximately 30,000 times, currently exceeding 450 trillion tokens on an annualized basis, and the scale of spending on the platform has also reached a remarkably impressive level. Meanwhile, the number of models supported by OpenRouter has increased from the initial four to over 500.

Figure: Growth of OpenRouter Token usage from establishment to acquisition

Menlo Ventures is fortunate to have participated in this journey. In March 2025, we participated in OpenRouter's seed round financing through the Anthology Fund established in collaboration with Anthropic.

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

In May 2025, we led OpenRouter's Series A financing, with Matt joining the company's board and Deedy serving as a board observer. Earlier this year, seeing the rapid growth in OpenRouter's customer base and revenue, as well as the company's product roadmap focused on building stronger "model intelligence" capabilities around model selection and evaluation, we further increased our investment in the Series B financing.

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

And just a few weeks ago, this happened: from Ramp to Cursor, over 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.

Figure: Group photo when deciding to lead OpenRouter's Series A

At first glance, Stripe does not seem to be the most natural acquirer for OpenRouter, but the two companies are surprisingly similar.

Both simplify what was originally a complex transaction process through a directly accessible API and charge a certain percentage fee for it. The only difference is that OpenRouter handles AI models.

According to Stripe's consistent statement, after the merger, both companies will still be doing the same thing: increasing "internet GDP."

In fact, over a year ago, OpenRouter already referred to itself as "the Stripe of the LLM space."

Core Value of OpenRouter

OpenRouter is one of the first companies Deedy engaged with after joining Menlo in 2024. This company is almost exactly at the core of our AI infrastructure investment logic.

Menlo proposed two judgments necessary for investing in OpenRouter in the "2024 Enterprise AI Report": AI spending will grow significantly, and developers will not use just one model but will adopt multiple models simultaneously.

Figure: The initial contact email Menlo sent to OpenRouter

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

For example, when you are just performing a simple NLP task, such as recognizing entities from text, you do not necessarily need to call the most cutting-edge and powerful models like Fable.

But the problem is that if users need to go to each model company's official website, register an account, create an API Key, securely store the key, adapt to the slightly different API specifications of each company, and finally manage all models themselves, the entire process becomes very cumbersome.

A unified model gateway sounds simple, but in reality, it is a much more challenging infrastructure problem than it appears. Few people are truly willing to build and maintain such a system long-term.

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

The more users there are, the better OpenRouter can predict model demand and handle larger loads; it also becomes easier to sign larger contracts with model labs, thus obtaining a more stable supply and demand for tokens.

Ultimately, this will create a cycle: new model labs will also want to prioritize landing on OpenRouter to gain distribution channels.

We have also observed another trend.

With the popularization of vibe coding, the number of software startups has rapidly increased. For a product aiming to enter the enterprise market, those that have already won the recognition of internal developers are often the ones that can ultimately be procured by enterprises.

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

OpenRouter is no different.

Since our investment:

• The number of tokens currently processed by OpenRouter has reached 30,000 times that of its early launch, maintaining about 33% monthly growth over the past three years, doubling approximately every 11 weeks.

• The model market has also rapidly expanded. 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. OpenRouter has now integrated over 500 models from more than 80 model providers, serving around 10 million users.

• Many important new models prioritize landing on OpenRouter, including models from OpenAI, X, and Meta. Mark Zuckerberg, who usually rarely tweets and even less often promotes other products, specifically announced the launch of Muse Spark on OpenRouter, as did Elon Musk. OpenAI also offers exclusive discounts for 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 have seen. Enterprises can use this product to uniformly allocate model resources, control access permissions, and manage internal AI budgets.

• Since OpenRouter can negotiate contracts uniformly with different model providers, it can offer extremely high service availability even when facing cutting-edge models.

How Did OpenRouter Get to Today?

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

OpenRouter is no exception.

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

The problem they initially wanted to solve was to avoid users being locked into a single model vendor while not having to provide their API Keys to different applications for using different models.

The design inspiration for this product came from cryptocurrency wallets, which is also related to Alex's previous experience founding OpenSea. At that time, Window supported a total of four models.

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

About a month later, they integrated the first batch of Anthropic v1 models, began automatically assigning requests to the appropriate models based on prompts, and created the model leaderboard that later became widely known, starting to use the name OpenRouter.

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

At that time, OpenRouter processed about 3 billion tokens per week. The models on the leaderboard then were completely different from the models familiar to everyone today.

Subsequently, the team launched Playground, allowing users to receive responses from multiple models through a chat interface.

By November of that year, OpenRouter had already supported 52 models, integrated over 2,000 applications, and processed about 8 billion tokens per week.

At this point, they had truly found PMF.

The Future of Model Routing

Contrary to many people's understanding, OpenRouter's core product is not simply "routing tokens for users," but rather becoming the most user-friendly AI gateway.

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

Recently, the rapid growth of enterprise spending on LLMs has become a real issue for companies like Uber, Coinbase, and Microsoft.

Therefore, "model routers" as a cost-reduction solution sound very appealing.

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

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

But the problem is that simply choosing models based on prompts is not a particularly effective method.

In agent scenarios, a task may need to run for a long time. To determine which model a request should be assigned to, a large amount of context needs to be understood.

For example, a simple instruction: "Find this file in the codebase."

It may only require a cheap, simple LLM, or it may need to call the most cutting-edge model. It depends on how large the codebase is and what context has been accumulated from previous tasks.

In multi-step agent tasks, if the router selects the wrong model at any step, the cost of that error will amplify along subsequent steps, ultimately leading to a noticeable decline in the quality of the overall task result.

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

It is these users that have allowed OpenRouter to gradually accumulate an extremely large dataset, which includes what prompts users submitted, which models were ultimately used for those prompts, what context was present during task execution, and what results were ultimately achieved, all while being largely unnoticed by the outside world.

This is where model routing becomes truly important.

In real production environments, OpenRouter can help enterprises control costs through more reasonable model routing while maintaining effectiveness as much as possible.

In the future, developers will not need to build complex evaluation systems themselves or continuously modify prompts for different models.

You may only need to log into the backend and see a prompt like this: "There are some scenarios in your codebase primarily used for summarization. If you switch from GPT 5.6 Sol to Muse Spark, you can save $100,000 a year. We have already completed the relevant evaluations for you automatically."

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 win over developers and then gradually enter the large enterprise market. Both have very simple and direct product design languages.

Andrej Karpathy once referred to OpenRouter as the "switch" for AI. What Stripe does is essentially similar: it has become that "switch" in the payment processing system.

This acquisition is also one of the earliest large transactions in the infrastructure field of the AI era, and it will not be the last. Infrastructure for managing models, costs, and computing power is gradually taking shape. This infrastructure is accelerating the formation of a new generation of giant companies faster than in the past.