
AI illustration of network hardware, representing routing between AI model providers.
Perplexity has launched its Orchestrator API for AI model access. Apps can use one link and one bill. If a request fails, the service can try another host of the same model.
The November 17 launch brings open-weight and closed models together. The firm hosts some open models. Other firms supply the closed models.
Connect an app to Orchestrator API
The quickstart guide lists three request formats. They are OpenAI Chat Completions, OpenAI Responses and Anthropic Messages.
Users change the base URL and API key in their client. They then pick a model from the list. The model sets which features they can use.
The Orchestrator API keeps no chat state on its servers. Apps must send the chat history needed for each request. Our Perplexity guide maps its search, agent and code tools.
A retry keeps the same model
The technical launch post explains how it checks each host. It looks at faults, load and speed. A retry uses the same model at a new host.
The firm reports higher success rates for Claude and GPT requests through its service. It compared hosts from October 4 to November 2. It left out client errors such as invalid requests.
The routing guide sets out the error rules. Bad requests get no host retry. If an output stream breaks, the text is partial. Users pay for tokens sent before it broke.
Each model has its own token rates. The host used for a request does not change the rate.
A preview can choose the model
The Orchestrator API offers a preview that can choose models. Teams can opt in to use it. They describe which tasks each model should handle. A model reads the chat and picks a route.
A default model takes work when no route fits. It also takes over if the choice step fails.
Our local AI versus cloud AI guide covers the wider hosting choice. Apps still manage their tools and chat history. They also check the quality of the work.
Read also
Popular news

Nvidia revenue doubles to $115.8 billion as Rubin ships

ChatGPT Can Bypass the “I’m Not a Robot” Test: How Is That Possible?

When GTA 5 launched, ChatGPT was still nine years away

Mistral launches Vibe Desktop for local and cloud coding tasks

Anthropic IPO priced at $83: Claude’s Nasdaq debut explained

