Perplexity Puts Open-Weight Models Behind One API, With OpenAI and Anthropic Formats

Perplexity has put its Router API into private preview, offering one API key for open-weight models it hosts. Developers choose a catalog model through familiar OpenAI or Anthropic request formats, while Perplexity says it sends each request to a healthy deployment.

One catalog, three request languages

The Router accepts OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages formats. Perplexity says any catalog model can be called through any of those schemas, regardless of its original provider. The service is therefore a compatibility layer around Perplexity-hosted models, rather than a new request protocol for developers to adopt.

For an existing OpenAI integration, Perplexity says the required changes are the base URL and API key. Anthropic SDK users point to a different Router base URL because that SDK adds its own versioned Messages path. Both paths use Perplexity credentials, removing the need for separate provider accounts and SDKs.

Model selection is not automatic

The routing is about availability, not task-level model choice. A developer still names a model from the catalog; Perplexity then routes that request to a healthy deployment. The documented catalog includes Perplexity IDs for DeepSeek V4 Flash, GLM-5.2, and Kimi K3.

That split also applies to billing. Perplexity says customers pay the selected model’s published rate even if another deployment serves the request. Its catalog endpoint lists current model IDs and base input, output, and cache-read prices in U.S. dollars per one million tokens.

Familiar endpoints, some Router-specific details

The documented chat-completions endpoint uses bearer-token authorization and the OpenAI conversation schema. It supports function tools with a name, parameters, and strictness setting. Streaming uses server-sent events, and an include_usage option returns token usage in the final streamed chunk.

Applications using the OpenAI Responses format face one important difference: Perplexity documents that endpoint as stateless, requiring the full conversation in every request. Router is meant for direct model access with a developer’s own prompts and tools; Perplexity directs users wanting web-grounded answers with built-in citations to its separate Agent API.

Access remains the immediate limit

Router API is not yet broadly available. Perplexity labels it a private preview and asks developers to request access by email, making the current release a gated test of its compatibility and managed-routing approach.