AI NewsDev toolsAnnouncement

Perplexity launches Orchestrator, one API for 85 models and 12 providers

Perplexity Orchestrator is one API that reaches 85 models from a dozen providers. Each model keeps its own token price. There is no extra platform fee.

AI News

Editorial2 min read

LinkedInX
Reveneau illustration of a model routing API

Why it mattersTeams using Claude, GPT and open models today keep three accounts and three invoices. Orchestrator reduces that to one key and one bill.

A small team that uses OpenAI, Anthropic, Google and an open-model host today keeps four accounts. Four API keys. Four invoices. Four budget caps. Perplexity launched Orchestrator on 17 November to collapse that stack. The service is live.

Orchestrator is one HTTPS endpoint. Perplexity says it reaches "85+ models and more than a dozen providers". The list includes OpenAI, Anthropic, Google and xAI. It also covers open-weight models such as GLM, Kimi and DeepSeek. The API accepts three request formats: OpenAI Chat Completions, OpenAI Responses and Anthropic Messages. A team moving between those changes the endpoint and keeps the request body.

What changes for a buyer

Billing is the real change. Perplexity writes that each request "counts only against the answering model's token usage". There is "no additional platform fee on top of credits". A call to Claude Opus 5.5 through Orchestrator pays Anthropic's published Opus 5.5 rate. A call to GPT 6.1 Sol pays OpenAI's rate. An admin gets one invoice. The dashboard reports spend by model, project, API key and date. An admin sets caps.

Perplexity also reports a reliability number from its own traffic. The window was 4 October to 2 November. The company writes that "Claude and GPT requests through Perplexity had higher success rates than any individual provider measured". The method is a retry. When the first provider errors, rate-limits or times out, Orchestrator retries the same model on another deployment. The number is Perplexity's own measurement. A team adopting it should repeat the test on its own traffic.

Two pieces still in preview

Routes and Auto are in preview on launch day. They are not generally available. Routes lets a team write plain-English task descriptions and pair each with a model. The team then calls perplexity/auto. Perplexity's decision model picks which model answers. Small classification jobs can go to a small model. A large model answers only when the task requires one. The team owns the route descriptions. The reports show which model served which request.

What a buyer should check

One API is also one more vendor between a product and the model it depends on. Map which features Orchestrator exposes today on each provider: batch pricing, prompt caching, tool use, streaming, image inputs. A missing feature is a reason to keep a second key for that workload.

A team already running a routing layer in house gains something new. Orchestrator is the first vendor that bundles one bill with the routing itself. A team paying a router for switching and each model vendor separately has one answer that covers both. The trade is a second company that must be running for the product to work. That is the question every "one API" pitch raises.

Source

Perplexity, with a companion design post on how the routing works.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project