OpenRouter: from Seed to Stripe — with OpenRouter’s Alex Atallah & AMP’s Anjney Midha

The conversation opens with Alex Atallah explaining a product principle he wrote about in early 2023: “PubSub as a product principle.” He sees products as an intersection between publishing data and subscribing to it. In a marketplace, suppliers publish products to a SKU, while consumers subscribe to that SKU when they want to buy. Human consumption is episodic—people focus on a product only while purchasing it—but agents and inference consumers use models continuously and change which models they consume from over time. Alex describes OpenRouter as a combination of a conventional API and a marketplace. It creates model “slugs,” offers an auto-router, and provides model and provider options that developers can subscribe to while continuously making routing and purchasing decisions. The host notes that, when OpenRouter began, it was not yet obvious that developers would want to swap models frequently rather than use native SDKs. Alex’s answer is that the ecosystem needed a neutral home where developers could discover models, compare usage, and use different providers without rebuilding their integrations each time.

Alex traces his conviction to the end of 2022 and early 2023. At that point, OpenAI was essentially the only major game in town, alongside Cohere and early open-weight attempts. When Llama appeared in January 2023, it was exciting because it outperformed GPT-3 on one or two benchmarks, although Alex felt it was not yet an engaging chat model. The Stanford Alpaca project changed his view: the team generated synthetic data, fine-tuned Llama, and produced a roughly 7- or 13-billion-parameter model for only about $600. Alex recalls using it on an airplane and finding that, in many cases, he could not distinguish its answers from ChatGPT’s. He concluded that valuable data could be “pressed into” a model and repackaged as a service, creating a new business model for data owners. Small teams could imitate or specialize frontier capabilities rather than waiting for a single lab to make every product decision. But that would require a marketplace for discovering the resulting models and services. Hugging Face was the closest existing home, yet Alex says it did not initially provide the same combination of closed- and open-source models, usage information, and inference access that he wanted.

Anjney Midha describes meeting Alex around 2011 or 2012 through Stanford Review, a libertarian student newspaper founded by Peter Thiel. They both supported adding a technology section to a publication focused mainly on politics, particularly because debates such as net neutrality combined technical and political questions. Their professional paths later overlapped through Discord, where Anjney ran the platform during the crypto and NFT boom and had to deal with security problems, phishing, social engineering, DDoS attacks, and billions of dollars in NFT-related activity involving OpenSea. Discord’s early access to GPT-3.5 led to two major AI experiments: Clyde, a first-party assistant, and LLM-based content moderation. The moderation use case failed in an important way: OpenAI’s post-training guardrails caused the model to refuse prompts that Discord needed it to handle. A Harry Potter fan server could trigger refusals because of trademarked content, and a detective-writing community could be blocked from drafting a violent chapter. Discord had 250 million monthly active users and more than 5,000 people working globally on moderation, so Anjney wanted access to weights or at least more control. OpenAI declined, convincing him that enterprises would need open models plus a management and orchestration layer. When Llama and later Mistral arrived, he recognized that Alex’s emerging marketplace could become that layer.

Alex says the biggest objection investors raised was “big model win”: scaling laws and network effects would supposedly cause one company to dominate AI just as Google dominated search. He and Anjney reject that analogy because LLMs are not only interfaces; they are infrastructure on which many new businesses can be built. A Google-like monopoly over a system on which the whole economy depends would be far more consequential, while the economics of producing competing models are more decentralized. Anjney’s reasoning came from the opposite direction. As an investor in Anthropic, Mistral, Black Forest Labs, and Luma, he saw multiple model teams producing innovative capabilities. Their common weakness was what happened after training: research teams often treated a finished checkpoint as the end of the work and simply exposed it as an API. He says Anthropic’s first Claude checkpoint sat internally for about a year, and when it was finally released there was no real distribution plan; the launch post relied on examples from friends and partners such as Discord, Juni Learning, and Notion. OpenRouter supplied the missing plumbing—key management, provisioning, endpoint management, versioning, and developer discovery—as well as neutral marketing and packaging. Alex adds that models are black boxes whose capabilities cannot be fully represented by a feature list. Developers need a neutral party to “shine lights” into the dark room of model options, show how models are used, and help them select and route among them. The founders push back on the description of OpenRouter as merely a “wrapper”: coordinating multiple APIs reliably, building the community, and putting a newly launched model in front of developers are strategic work, not a thin technical layer.

Alex explains that OpenRouter’s early growth followed lessons from OpenSea and its work with communities such as Axie Infinity. OpenSea repeatedly entered existing communities, learned what users actually needed, built the missing marketplace experience, and watched for signals such as users sharing OpenSea links. In AI, the early gap was developers encountering refusals, an inability to customize models, or no straightforward way to use open weights. Those blocked communities were the most valuable places to listen and build for. Anjney recalls a Discord and OpenSea integration call that shaped his view of Alex. The technically simple plan was to send users out of Discord to OpenSea, but Alex objected that this was a poor user experience and proposed embedding checkout inside Discord. Anjney says Alex prioritized the user even though the embedded approach could be seen as adversarial to OpenSea’s interests. He viewed that combination—technical understanding, developer empathy, and end-user judgment—as unusually valuable. The crypto ecosystem also served as a rehearsal for generative AI. Discord hosted early communities and applications, including Axie Infinity, Midjourney, and an internal bot using OpenAI models. Midjourney’s Discord server let users see other people’s prompts and copy them, solving the paralysis of an empty prompt field; the service grew from launch to a $100 million revenue run rate in less than eight months, according to Anjney. Stable Diffusion then made it possible for developers to build their own Midjourney-like products, creating demand for an API and a neutral distribution layer such as OpenRouter.

The first clear proof that OpenRouter’s provider marketplace added value came with Mistral’s Mixtral 8x7B. The model triggered claims that an open-weight model was the best in the world and appeared to outperform GPT-4 on some examples. Alex notes that some of the perception may have come from speed: Mixtral’s mixture-of-experts architecture was extremely efficient, and people can interpret a faster answer as a smarter one. In practice, developers could ask a fast model first and manually switch to a stronger or more expensive model when necessary—the beginning of a human routing process that OpenRouter later automated. OpenRouter cleaned up a messy inference landscape and allowed providers to compete on price, giving developers the best available price in one place. Anjney recalls Mistral’s Guillaume Lample being more modest about the model’s quality, but the launch still became an important demonstration of competition among inference providers. The team also discusses an early “mixture of models” prototype from early 2024. It asked several models for answers, then used another model to fuse those results and displayed the intermediate outputs in a large Kanban-like interface. The idea was directionally right, but the technology and form factor were immature: the fused answer was often no better than the strongest individual model. OpenRouter deleted the product rather than letting it distract the team. Later, after the top models became closer in quality, Alex tested fusion on architecture plans and found that the models repeatedly judged the fused result better than their individual answers, leading the team to revisit the idea.

The discussion separates OpenRouter from Arena, another project in which Anjney invested and initially served as CEO while helping its Berkeley founders spin out. Arena’s original mission was to make model evaluation more statistically reliable for research and post-training teams, using methods developed by its founders as Berkeley scientists. OpenRouter’s target was different: developers who take research results and turn them into applications deployed to the world. Anjney says the two companies might look adjacent from a distance, but their customers, missions, roadmaps, and data policies were distinct. OpenRouter does not see prompts or completions by default; organizations must opt in if they want that data exposed. That distinction leads to a broader argument for focus. Alex says founders naturally want to own every adjacent layer, but serving multiple customer types makes it hard to build a product that exceeds expectations. Anjney agrees that focus helps both the product and the market understand what a company is known for. He uses Anthropic as an example: from day one its mission was to responsibly commercialize AI as a “bare programmer,” and it prioritized coding despite the attention going to image and video models. The company’s early model releases emphasized capabilities such as writing and long context, but the core research direction remained coding, long-horizon tasks, and agentic programming. OpenRouter declined to launch a fine-tuning-as-a-service product in which users could submit a few YouTube videos and receive a model that spoke like the people in them. The prototype was technically interesting but received little use, and the team believed the experience was closer to retrieval over source videos than a genuinely differentiated model. Rather than expand into every possible infrastructure service, OpenRouter chose to remain a neutral marketplace and partner with companies specializing in fine-tuning and custom models.

Alex describes OpenRouter’s growth as a series of inflections tied to model and application launches. Through May 2024, the platform was heavily oriented toward prompts and limited coding use cases. Claude Sonnet 3.5 produced a major improvement in coding, causing a surge in developers and application volume and making users more attentive to inference costs. Mixtral then supplied a cheaper open-weight alternative, creating a recurring pattern: a frontier model caused usage to surge, developers examined their bills, and open-weight models later provided more cost-efficient options. By the end of 2025, OpenClaw introduced a different user and architecture. It attracted productivity users and internet creators, not just developers, and sent periodic “heartbeat” requests to check whether the agent was alive. OpenRouter’s auto-router was valuable because users did not want to pay for an expensive model every 30 minutes. Hermes followed with an emphasis on community, memory, and skills. Alex says OpenRouter’s leaderboard and app rankings effectively form a historical record of changing AI usage, from coding agents and terminal tools to newer agent products. He also says the platform has grown past 10 million developers and is now handling token volume at an accelerating rate. The Stripe partnership is framed primarily as a trust-and-safety and distribution partnership. OpenRouter blocks stolen-card fraud, account compromises, unauthorized inference resale, runaway agents, and other abuse; Alex says the company blocked ten times as much dollar volume in one month as the previous month. Anjney argues that tokens are becoming a new unit of value, much as online payments did, and will require security infrastructure comparable to Stripe Radar. He expects AI agents to automate abusive behavior at much greater scale. OpenRouter will keep its name, brand, product, and roadmap, while the near-term goal is to move faster with Stripe and make it easier to accept tokens, bring third-party inference, and build on top of models.