OpenRouter
An AI inference gateway and marketplace that provides one API for models served by multiple underlying providers.
OpenRouter is an American artificial-intelligence company that operates an inference gateway and marketplace. Its service gives developers access to models from multiple providers through a common API, account, and billing system. OpenRouter routes each request to an eligible provider, which performs the model computation.12
The company is an inference provider at the user-facing service layer, but does not ordinarily operate as a cloud GPU provider. It supplies access to managed model endpoints rather than customer-administered accelerator instances. The developer of a model, the company hosting it, and OpenRouter can be separate organizations.
History
OpenRouter was founded in early 2023 by Alex Atallah and Louis Vichy. Atallah had previously co-founded the NFT marketplace OpenSea and served as its chief technology officer.13 The service was developed in response to the growing number of model vendors, each with separate accounts, interfaces, and availability characteristics.
On 28 May 2026, OpenRouter announced a US$113 million Series B financing led by CapitalG, with participation from Andreessen Horowitz, Menlo Ventures, Sequoia Capital, and other investors. The company said that its network contained more than 400 models and 70 providers at the time.4 These figures vary as models and endpoints enter or leave the catalog.
OpenRouter introduced a new visual identity in July 2026. The redesign uses a geometric “OR” mark inspired by Bauhaus forms.5
Service
Applications send requests to OpenRouter using a model identifier and an OpenRouter API key. The gateway authenticates and records the request, selects a provider endpoint, and relays the response. Its API follows many conventions of the OpenAI API, allowing existing client libraries to be used with a different base URL.2 Model-specific features, parameter support, and response formats can still vary.
The service distinguishes between a model and the provider endpoints that serve it. Several providers may offer the same open-weight model, but use different hardware, quantization, context limits, or inference software. OpenRouter publishes endpoint metadata and permits a request to be restricted to particular providers or deployment characteristics.
OpenRouter’s routing system can order providers by price, latency, or throughput and can fall back to another endpoint after a failure. Routing can also be restricted by quantization and data policy.6 Unless the caller specifies a provider, successive requests to the same model can be processed by different hosts.
Pricing
OpenRouter consolidates usage charges across providers. The listed token price generally reflects the price of the selected endpoint, while OpenRouter charges a fee when users purchase credits.7 It also supports bring-your-own-key access for some providers, allowing their own API credentials and billing relationship to be used through the OpenRouter interface.
This arrangement gives model hosts access to demand through a shared marketplace and allows applications to use several providers without separate integrations. It also makes OpenRouter an additional operational dependency between the application and the inference host.
Data handling
OpenRouter states that prompt and response content is not stored by default, although request metadata such as model, provider, token count, and latency is retained. Content logging can be enabled by the user for supported features.8
The policy of the selected provider applies separately. OpenRouter documents provider-level retention and training policies and allows routing to be limited to endpoints marked as supporting zero data retention.9 A request may consequently be subject to policies from both OpenRouter and the company performing the inference.
References
Footnotes
-
About OpenRouter, OpenRouter. ↩ ↩2
-
Frequently asked questions, OpenRouter documentation. ↩ ↩2
-
Investing in OpenRouter, Andreessen Horowitz, 26 June 2025. ↩
-
OpenRouter raises $113M to build the universal AI interface, OpenRouter, 28 May 2026. ↩
-
OpenRouter’s new look, OpenRouter, 13 July 2026. ↩
-
Provider routing, OpenRouter documentation. ↩
-
Data collection, OpenRouter documentation. ↩
-
Zero data retention and provider data policies, OpenRouter documentation. ↩