This past May, a Silicon Valley AI infrastructure startup raised $113 million from a growth investment fund under Google's parent company. Its valuation more than doubled in a year, reaching $1.3 billion. The company doesn't build AI models or distribute content. It acts as a router, connecting user requests across hundreds of AI models. What caught people's attention wasn't just the valuation figure. The more striking number was that API usage had grown fivefold over the previous six months. It was a signal that the judgment of which AI model to use, and when, was itself becoming a viable business.
What's Happening Across 200 Models
OpenRouter is a service that aggregates more than 200 AI models—including GPT-4o, Claude, Gemini, and Llama—under a single API interface. When a developer or company sends a request, the platform assesses the nature of the task, cost, and response speed, then routes it to the most appropriate model at that moment. If one model goes down, it automatically switches to another. Users can specify a model directly, or simply set priority criteria and let the platform decide.
CapitalG, which led the Series B round, is Alphabet's growth-stage investment fund. The investment is notable given that Google itself runs an AI ecosystem—through its Gemini models—that competes directly with OpenRouter. Here, a company that owns competing models simultaneously bet on multi-model routing infrastructure. It reads as a judgment that the connectivity layer between models may hold more durable value than the market-share battles among individual models themselves.
The company cites a simple driver for its growth: the recognition that no single model is optimal for every task is spreading quickly through the market, and demand for finding the right balance between cost and performance is building stronger than expected. The logic isn't hard to follow. Some models excel at document summarization; others are built for code generation. Open-source models can be significantly cheaper than commercial ones, even if their raw performance is somewhat lower. Matching the right model to the right task means either better results for the same cost, or the same results for far less. Demand for infrastructure that automates that matching drove usage up fivefold in six months.
The Counterarguments Behind the Growth Numbers
Behind the growth figures attracting investor attention, skeptics exist. OpenRouter's most realistic threat isn't a competing service—it's the model providers themselves. OpenAI, Anthropic, and Google each control their own API pricing and distribution channels. If they lower their fees or start bundling multi-model routing natively into their own platforms, the differentiation OpenRouter offers gets thinner. There have already been reports that some providers are experimenting with routing-like features within their own APIs.
The concept of "optimal routing" itself remains blurry. Whether to prioritize speed, cost, or accuracy varies by task—and even for the same task, context shifts the calculus. If users don't configure these criteria precisely, it becomes difficult to verify after the fact whether the routing decision the platform made was actually the best one. As scale grows, that opacity can accumulate into cost overruns or quality inconsistencies.
This is also a period when AI infrastructure valuations have risen broadly and rapidly. OpenRouter is demonstrating strong execution, but open-source alternatives offering similar functionality at significantly lower cost also exist. The fivefold usage increase is evidence of growth, but it doesn't automatically guarantee the durability of its market position.
What This Signals for Practitioners
What OpenRouter's growth reveals is that a mindset of approaching AI as a question of "which tool to use" is spreading through real-world workflows. If "should we use ChatGPT?" was the practical question in 2023, then "which model should we use for which task?" is the more pressing concern in 2025. The two questions look similar on the surface, but they reflect meaningfully different ways of thinking.
The first is a decision about whether to adopt a tool. The second is a judgment about how to deploy it. When there was only one tool, the first question was all there was. Now that there are dozens, the second has a more direct impact on actual costs and outcomes.
This hits close to home for solo operators and small teams. Even if you're paying for an AI subscription and using it for most of your work, cheaper models often deliver equivalent results for simple summarization, classification, or first-draft generation. Conversely, for contract review or in-depth report analysis, a model one tier up from your default subscription may produce meaningfully better output. Cost inefficiency tends to appear exactly where task and model are mismatched.
I want to make one thing clear. We're entering a phase where the ability to judge which AI tools to use is more directly tied to practical efficiency than knowing how to use AI at all. It's worth asking yourself: do you know which tasks your current AI is overcharging you for? Are you routing image analysis, document summarization, and brainstorming through the same single tool? Do your criteria for choosing tools come from habit or from deliberate judgment? Regardless of whether you adopt OpenRouter tomorrow, there's a real difference between being able to answer those questions and not.
Demand for outsourcing the judgment of which tools to use opened a $1.3 billion market. Trace that demand back to its source, and you can start to see where the people who make that judgment themselves are finding their savings.
