“Jagged intelligence” that’s how AI researcher Andrej Karpathy describes the uneven behavior of modern large language models (LLMs), where a system can solve Olympiad-level math problems in one moment and fail a basic logic test the next.
OpenRouter, a startup founded in April 2023 by Alex Atallah and Louis Vichy, has raised $40 million across seed and Series A rounds to help developers navigate that chaos. Its platform offers a single API to route prompts across hundreds of proprietary and open-source AI models, abstracting away the complexity of switching between vendors like OpenAI, Anthropic, Meta, and dozens of cloud GPU providers.
The seed round, led by Andreessen Horowitz in February, was followed by a $28 million Series A led by Menlo Ventures in April. Sequoia Capital and angel investors including Dylan Field, Tobi Lütke, and Ashton Kutcher also participated. The company is now valued at approximately $500 million, according to a person familiar with the deal. Atallah declined to comment on the valuation.
Atallah started the company after a simple but crucial question: Will inference become a winner-take-all market? In the early months of 2023, OpenAI dominated, but Atallah noticed growing developer frustration not just over cost, but over moderation policies and model behavior.
“In January, we saw the first signs of people wanting alternatives. They wanted more transparency, different moderation rules, and more control,” Atallah recalled. “That was the first crack in the idea of a single-model future.”
The shift gained momentum when Meta’s LLaMA 1, released in February 2023, outperformed GPT-3 on several benchmarks despite being small enough to run locally. In March, Stanford’s Alpaca project fine-tuned LLaMA on GPT-3 outputs for under $600 demonstrating that smaller, custom models could replicate the capabilities of far larger ones. That was the unlock.
“It meant you didn’t need a $10 million compute budget anymore,” Atallah said. “You could distill knowledge and style into smaller models. There needed to be a way to discover and use all of them.”
OpenRouter began as an aggregator of these emerging models. The team had previously built Window AI, a Chrome extension that let users inject their preferred LLM into any web app. That experiment laid the groundwork for OpenRouter, which has since evolved into a full marketplace with more than 450 models and 60+ GPU providers. Customers pay once and route between models without rewriting code or renegotiating contracts.
“Our business is a one-stop shop for all models,” said Vichy. “Whether you care about performance, price, latency, or safety, we route to the best option for your workload.”
The traction has been substantial. OpenRouter processed more than 8.4 trillion tokens in May alone, with customer inference spending hitting $8 million that month up tenfold from October 2024. The company earns roughly a 5% fee on inference costs. More than a million developers now use the platform.
Enterprise adoption is also accelerating. OpenRouter offers advanced features such as caching, smart model routing, failover for downtime, observability dashboards, role-based access controls, SOC 2 compliance, and support for customer-owned GPU clusters. Its platform helps users switch between providers with zero code changes while providing detailed latency and throughput data for each model.
“We think inference spending will eclipse salaries as the dominant operating expense for most knowledge-based companies over the next five to ten years,” said Chris Clark, OpenRouter’s COO. “And routing intelligently will be non-negotiable.”
That complexity is already surfacing. As more providers host the same open-source models each with different features, rates, and behaviors, developers need more than just access. They need control.
“There’s a model explosion, but also a provider explosion,” said Atallah. “Some support tool calling, some don’t. Some have better uptime. Some have broken streaming. We unify that.”
The team has also built AI-native middleware that lets any model be enhanced with plug-ins, for example, real-time web search or PDF parsing streamed directly into the output. The goal is to create feature parity across the model landscape so that developers can build once and route flexibly.
With new modalities on the horizon, such as transfusion models that combine language and image generation, OpenRouter is preparing for an even more fragmented future. The roadmap includes regional GPU routing, fine-grained model categorization (e.g., best model for converting Japanese to Python), and deeper observability into prompt behavior and cost optimization.
Matt Murphy, general partner at Menlo Ventures and a board member at OpenRouter, believes the company’s long-term edge lies in the data.
“They’re not just connecting to models—they’re collecting intelligence about intelligence,” Murphy said. “I’m most excited about the future revenue model because the data asset they have will be near impossible for anyone to replicate.”
After a year of exponential growth, Atallah’s original question, Will AI inference be winner-take-all? has a clear answer.
“The future is multi-model,” he said. “And developers need the tools to keep up.”
