OpenRouter CEO Alex Atallah: Dedicated Models Favor New Labs

OpenRouter CEO and co-founder Alex Atallah stated that if the market begins to seek dedicated models optimized for specific scenarios and branded separately, there will be an interesting shift in the diversity of large models, which is beneficial for the market and will create space for new labs that are not on the cutting edge. This judgment is a directional statement and does not disclose new transaction or share numbers.

Atallah was previously the co-founder and CTO of OpenSea, and later, along with Louis Vichy, developed the browser plugin Window AI into a unified access point, which then evolved into the multi-model inference marketplace OpenRouter: allowing switching between multiple labs and inference providers based on price, latency, and availability with a single access. The company positions itself as "increasing the neural diversity of artificial intelligence," believing that no single model can handle all scenarios.

In May 2026, OpenRouter announced a $113 million Series B round, led by CapitalG, a subsidiary of Alphabet, with participation from Nvidia NVentures, ServiceNow, MongoDB, Snowflake, Databricks, Andreessen Horowitz, and Menlo Ventures. At that time, it was disclosed that the weekly processing volume increased from about 50 trillion tokens to about 250 trillion; subsequent public rankings showed continued expansion of the platform's weekly volume, with statistics indicating a twenty-fold increase within a year. Atallah mentioned in a podcast that the platform accounts for about 1.5% to 2% of global token traffic and that the usage of direct connections to OpenAI, Anthropic, Gemini's official interfaces, and bundled products like Cursor and Claude Code is systematically underestimated.

He provided the architectural judgment that cutting-edge orchestration models assign deterministic sub-tasks to cheaper open-weight sub-agents. In July 2026, approximately 70 new models were added to the platform in just one month, about one every 10 hours; the new supply comes from agent labs, with Cognition and Cursor having developed their own models. He specifically mentioned that small labs in the U.S., such as Poolside, are effective in coding small models. On the open-weight side, GLM, Kimi, and DeepSeek are seen as quality steps; he also stated that the U.S. is "very behind" in the quality of open-source weights.

The routing leaderboard and revenue leaderboard do not overlap. In terms of platform usage, DeepSeek, Zhiyu GLM, Tencent Hy, and Xiaomi MiMo have repeatedly ranked among the top daily active users, with statistics indicating that U.S. labs' share on OpenRouter has dropped from about 70% to about 30%; based on request share, OpenAI still ranks high on the weekly leaderboard. In terms of inference provider revenue, OpenAI, Anthropic, and cloud resellers still command higher prices. This means that "who is called more" and "who receives the money" have been separated.

In market mechanisms, buyers need to break down the same workflow into coding, retrieval, long context, and cheap drafts for developers and agent runtimes; sellers are new labs that cannot afford to compete with cutting-edge models but can establish a brand in a single benchmark or industry. Traffic passes through the routing layer, where price differences and fault switching determine who to call next. Beneficiaries are routing platforms and labs that can label "certain tasks as cheaper and sufficient"; those under pressure are generic models that rely solely on overall leaderboard IQ premiums but are downgraded to sporadic calls by the orchestration layer. Funding has shifted from "buying an all-purpose model" to "buying a row of dedicated models by task," with settlements occurring in token bills rather than model launch events.

Additionally, the platform has also featured anonymous cutting-edge and free window models to test price-performance curves; it also warns that stealth models may enable logging. The routing layer does not default to binding a single inference provider, and the same model name may fall under different hosting parties, meaning quality and compliance no longer equate to the lab brand itself.

Source: Public Information