OpenRouter: Building the Neutral Routing Layer for a Multi-Model Future

The conversation opens with a token-economy framing that runs through the whole episode. Anjney Midha and Alex Atallah both treat tokens as a new unit of value streaming across the internet, and they argue that the more valuable that flow becomes, the more bad actors will try to capture it. Midha compares this to online payments, which grew past a trillion dollars in roughly a decade and forced the creation of entirely new fraud defenses. He expects the token economy to reach around $5 trillion within five years and perhaps $10 trillion within ten. Atallah adds that OpenRouter blocked ten times as much dollar volume last month as the month before, and that the types of token fraud are diversifying.

Why multi-model was the bet

Atallah frames OpenRouter as a blend of a normal API experience and a marketplace. In his telling, human consumers buy discrete products in ad hoc bursts, while agents and inference consumers subscribe continuously and keep changing the SKUs they consume. That is why he describes OpenRouter through a pub/sub lens: suppliers publish models, and consumers subscribe to topics they consume whenever they want.

The origin story starts with the open-weights wave. At the end of 2022, OpenAI and Cohere were effectively the only games in town. Llama arrived in early 2023 and outperformed GPT-3 on a benchmark or two, but it was not an engaging chat model. Stanford’s Alpaca then showed that the gap could be closed cheaply: roughly $600 of synthetic data and fine-tuning produced a model that was hard to distinguish from ChatGPT in many cases. Atallah saw this as a new business model, where a small team can compress valuable data into a model and sell it. It also meant a whole ecosystem could emerge around a breakout app, and that ecosystem would need a marketplace to discover its products.

Hugging Face was the nearest analogue, but it lacked closed-source models and did not show who was using what, or why. That gap mattered to Atallah while he was trying to understand why people chose different emerging models.

Discord as the proving ground

Midha’s account is a useful counterweight, because his path to the multi-model thesis came through enterprise deployment. As VP of platform at Discord, he was responsible for deploying models to roughly 250 million monthly active users. OpenAI’s content moderation was too rigid for that job. Its post-training prompts were heavy-handed: fan communities discussing Harry Potter were refused, and detective stories with violent chapters were blocked. Discord wanted access to model weights so it could encode each server’s own norms into a custom moderation layer, but OpenAI replied that it was a closed-source company. Midha’s conclusion was that enterprises would need open models and, eventually, a control plane to orchestrate them. Atallah’s marketplace idea came from the same tension: when one model refuses, users want to switch to another.

The “one model wins” objection

The most common objection VCs raised was that scaling laws and network effects would consolidate the market around one company, a Google-style monopoly. Atallah argues that search was not as service-like as LLMs, which are both user interfaces and platforms for building new businesses, and that competitive economics here are more decentralizable. Midha reads the scaling laws as a feature rather than a bug. Having invested early in Anthropic and drawing on grad-school friends who believed in the bitter lesson, he saw OpenAI and Anthropic as two proof points that compute scaling works, and he treated a diverse ecosystem of models as the natural outcome.

His deeper argument concerns distribution. Research labs think in terms of capabilities but not in developer terms. When a checkpoint finishes, labs often put it out as an API and then hear crickets. Midha describes Anthropic’s first Claude launch, which had no real developer distribution plan; its example list was a Discord bot, a friend’s startup, and Notion. Keys, provisioning, endpoint management, and versioning were all missing. Google can push a new checkpoint to a billion devices overnight because it already owns the surfaces. Without a neutral layer, labs must build all of that plumbing themselves. Midha says Anthropic took more than twelve months to reach its first $10 million in revenue. By contrast, OpenRouter could send a million developers to Black Forest Labs on launch day. Atallah adds that models behave as black boxes, even open-weight ones, so a neutral third party must shine light on what makes each model good, and that routing and discovery become central to go-to-market for providers and labs alike.

“Just a wrapper”

Midha is openly frustrated by the VC pushback. Some investors dismissed OpenRouter as “just a marketplace,” “a thin layer,” or “a wrapper” around other people’s APIs. Midha argues that orchestrating even three APIs in production, with the engineering and community design that requires, does not happen by default. He says Alex thought in systems and flywheels, something most ML researchers do not, since they still think in linear pipelines of pre-training, mid-training, and post-training. He invested without trying to reeducate the other VCs, and Menlo later marked the company up tenfold. He calls the “it’s just a rapper” framing reductive, and the two agree that everything is a wrapper in some sense.

Community bootstrapping and product taste

The bootstrap story runs through Discord and the Axie Infinity community, one of many communities the team used to find real problems. Their rule was to build things the community actually wants, and users sharing links was a clear signal of importance. Midha recalls a Discord and OpenRouter integration in which Atallah rejected the plan to bounce users out to OpenRouter’s site. Atallah pushed for an embedded iframe that kept users inside Discord, even though that cut against OpenRouter’s own interest. Midha says the moment showed a rare pairing of developer-flow understanding and user-experience priority.

Crypto, Midjourney, and the early generative ecosystem

Midha argues that crypto served as a dress rehearsal for generative AI. Discord’s NFT and crypto communities ran at scale and forced the infrastructure work that later supported AI apps. Midjourney started as a small Discord bot built by David Holz, and the team helped make the Discord server the primary interface, because a single-player web app showed users a blank canvas and produced poor retention. Watching others prompt in public was what made the product usable. Midjourney reached roughly $100 million in run-rate revenue within eight months of launch. When Stable Diffusion arrived, developers could build their own Midjourney, but they needed an API to do it without hosting weights, and that need helped shape OpenRouter’s purpose. Atallah’s first project was a Chrome extension called window.ai, built on Plasmo, which let web pages call a user-chosen model. Its lesson was that discovery requires more space and visuals than a browser extension can offer, and that the product must work for both humans and agents.

The pair also discuss why Discord is not OpenRouter’s main surface. Midjourney is an end-user application, while OpenRouter needs codebase integration, governance, and data policies that a chat server cannot provide.

Mistral and the inference marketplace

Mistral’s Mixtral 8x7B release in late 2023 triggered a price war. Its 8x7B model was perceived as better than GPT-4 partly because it was fast. Midha recalls Guillaume Lample saying it was “an okay model, not that good,” and speculates that speed made people perceive intelligence. OpenRouter cleaned up a messy inference landscape by letting providers compete on price, which gave users the best price in one place. Atallah calls this the first clear example of a provider marketplace adding value for developers. The discussion also turns to a “humans as router” pattern, where a person tries a fast model first and upgrades manually, which later became OpenRouter’s auto-router.

OpenRouter versus LM Arena

Both Midha and Atallah stress that the missions differ. Midha was LM Arena’s first CEO for a few months, when the project was an evaluation service that came out of Berkeley research on statistical correction of eval estimates. Its ideal customer was a post-training researcher at a lab. OpenRouter’s ideal customer is the developer who turns research into a deployed application. Atallah notes that OpenRouter cannot see prompts or completions unless an organization opts in, and that LM Arena’s model relies on data and labs in a way that makes it hard for one company to do both. He says the two could not easily be combined without confusing their brands.

Focus

Atallah argues that focus is underrated even in the age of AI, not only because it produces a better product but because the market can map a brand to a problem. Midha ties this to Anthropic’s early mission of responsibly commercializing AI pair programming, which set aside more glamorous detours such as image and video models. He says the coding evals have been the company’s main metrics from the start. OpenRouter’s own detours included a fine-tuning-as-a-service prototype that turned YouTube videos into a model imitating the speaker. It was never widely used, partly because the marketplace took priority. Atallah notes that fine-tuning as a router service, the kind that could export traffic from frontier labs, has not yet materialized, and that the company chose to partner with inference providers rather than compete with them.

Mixture of Models and fusion

In early 2024, OpenRouter launched MoM, or Mixture of Models. The product let users choose several models, fused their outputs with a final model, and displayed the intermediate results in a visual board. Atallah says the technology was too early: the fused answer was sometimes worse than the strongest model and sometimes no better than it. The team deleted the code. Years later, with several frontier models performing well, Atallah ran a personal experiment in which he asked each model whether the fused architecture plan beat its own answer, and every model said yes. That led to a benchmarked Fusion product. He attributes the improvement partly to RL expanding the diversity of reasoning within each lab.

Growth and milestones

Atallah describes growth as moving in step with model launches. Through mid-2024, the traffic was heavy on roleplay and creative apps, with Dream Tavern among the top apps. Claude 3.5 Sonnet in mid-2024 caused a leap in coding volume and made developers look hard at their invoices, prompting a turn to cost-efficient alternatives. Mixtral 8x7B and then open-weight models offered those options, creating a cycle: frontier launch, usage surge, invoices, then cheaper open models. The leaderboard by coding agent gained prominence, with Cline near the top at one point. By late 2025, OpenClaw introduced a new user type, a productivity or creator user rather than only developers. Its heartbeat architecture repeatedly called models to check whether the agent was alive, so auto-routing made the cost of those calls manageable. Hermes followed and built around skill management and memory. Atallah says the leaderboard works as a movie of the AI space over time, and Midha notes that Andrej Karpathy said he no longer read LocalLLaMA and instead checked the OpenRouter leaderboard.

Why Stripe, and why fraud matters

The Stripe partnership grew from years of collaboration on abuse and fraud. Atallah describes token fraud diversifying into stolen cards, resold traffic against terms of service, compromised accounts, resold inference, and accidental runaway agents. He argues that gateways and general-purpose inference sellers are prime targets, while discrete intelligence products are less exposed. He expects pricing to move toward per-task or per-event models, with consumer apps favoring simple subscriptions, and he expects labs and providers to push users to commit and then bring their own inference elsewhere.

Midha ties this to Midjourney’s free trial. A user in China resold free subscriptions, and the trial was shut down. He sees that as an early version of the core problem: tokens are valuable, and wherever value scales, bad actors follow. He draws a parallel to Stripe, which accepted fraud as a customer-acquisition cost early on and built Radar into a security company. He argues that OpenRouter’s future depends on defending the token economy, and that agentic fraud, driven by autonomous agents rather than humans, will be the next and much larger wave. He is blunt that current labs can see only the agent behavior they train themselves, which is a small fraction of the problem, and that a shared defense layer across providers will be needed.

What comes next

Atallah says the Stripe deal accelerates go-to-market, moves OpenRouter upmarket faster, and pairs trust and safety with the developer experience. The OpenRouter brand, product, roadmap, and name will remain the same. In the next six months, the company expects to do what it would have done as an independent company, only faster. Midha closes the conversation by framing the work as building new sheriffs for a token economy that is not yet secure enough to trust.

• Tags:
• Open-Weight Models
• Stripe
• A16z
• Discord
• Token Economy
• Midjourney
• OpenRouter
• LLM Routing
• Model Marketplace
• Inference Gateway
• Model Fusion
• Agentic Fraud