This May, a Silicon Valley AI infrastructure startup raised $113 million from a growth 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 of its own, and it doesn't distribute content. It acts as a router, connecting user requests across hundreds of AI models. For many observers, the valuation wasn't the headline. The number that drew more attention was this: API usage had grown fivefold in the past six months. It was a signal that the judgment of which AI model to use, and when, is becoming a business in its own right.

What Happens Among 200 Models

OpenRouter bundles more than 200 AI models—including GPT-4o, Claude, Gemini, and Llama—behind a single API. When a developer or company sends a request, the platform weighs the nature of the task, cost, and response speed, then routes it to the model best suited for that moment. If a particular model goes down, traffic automatically shifts to another. Users can specify a model directly, or simply set their priorities and delegate the decision to the platform.

CapitalG, which led this Series B round, is Alphabet's growth-stage investment fund. Given that Google itself runs a competing AI ecosystem through its Gemini models, the investment is a curious one. The side that owns a competing model placed a bet on multi-model routing infrastructure at the same time. It reads as a judgment that the connective tissue between models may hold steadier value than any single model's fight for market share.

The company's own explanation for its growth is simple. The recognition that no single model is optimal for every task is spreading quickly through the market, and demand for finding the sweet spot between cost and performance is forming faster than expected. Why this market exists isn't hard to see, either. Some models excel at summarizing documents; entirely different ones excel at writing code. Open-source models, even where they underperform commercial ones, cost dramatically less. Match models to task types and you get better results for the same money, or the same results for far less. Demand for infrastructure that makes that call automatically drove usage up fivefold in six months.

The Skepticism Behind the Growth Numbers

Behind the growth figures investors are celebrating, there is also a skeptical view. OpenRouter's most realistic threat isn't a rival service—it's the model providers themselves. OpenAI, Anthropic, and Google each control their own API pricing and distribution channels directly. If they cut API rates or start building multi-model routing into their own platforms by default, the case for OpenRouter's differentiation thins out. There have already been reports that some providers are experimenting with routing-like features inside their own APIs.

The very concept of "optimal routing" remains fuzzy, too. Whether to prioritize speed, cost, or accuracy varies by task, and even the same task can shift with context. Unless users configure these criteria in fine detail, it's hard to verify after the fact whether the platform's routing decision was actually the best one. As scale grows, that opacity can surface as accumulated costs or inconsistent quality.

This has also been a period when AI infrastructure valuations have risen quickly across the board. OpenRouter is clearly executing well, but open-source alternatives offer similar functionality at a fraction of the cost. A fivefold jump in usage is evidence of growth—it doesn't automatically guarantee that the company's market position will last.

The Signal for Everyday Work

What OpenRouter's growth reveals is that a particular way of looking at AI—as a question of which one to use—is spreading through working life. If the practical question in 2023 was "Should we use ChatGPT?", the more practical question in 2025 is "Which model should we use for which task?" The two questions look similar on the surface, but they reflect quite different ways of thinking.

The first is a decision about whether to adopt a tool. The second is a judgment about how to deploy tools. When there was only one tool, the first question was the whole game. Now that there are dozens, the second question bears more directly on real costs and real outcomes.

This lands directly on solo entrepreneurs and small teams. Even if you pay a monthly AI subscription and use it for most of your work, a far cheaper model often produces equivalent results for simple summarization, classification, and first drafts. Conversely, for contract review or analyzing long reports, a model one tier above your default subscription may deliver better results. Cost inefficiency lives wherever the task and the model don't match.

Here is the one thing I want to pin down. We're entering a phase where the ability to judge which AI tool to use connects more closely to practical efficiency than knowing how to use any given one. Do you know which of your tasks your current AI is overkill—and overpriced—for? Are you running image analysis, document summaries, and brainstorming all through the same tool? Is your choice of tool coming from habit, or from judgment? Quite apart from whether you adopt OpenRouter tomorrow, there is a real difference between being able to answer these questions and not.

The demand to outsource the judgment of choosing tools opened up a $1.3 billion market. Trace that demand back to its source, and you can guess where the people who can make that judgment themselves are already saving money.