Perplexity announced the Orchestrator API on November 17, 2026, bringing proprietary and open-weight models under one API with consolidated billing. Its provider failover can retry another deployment of the model already chosen; a separate, optional preview can select a model for each request. Perplexity’s announcement and routing documentation describe the two functions.

How Orchestrator’s two routing functions differ

Provider failover changes the deployment, not the selected model. Orchestrator chooses among deployments of that model based on observed error rates, capacity and latency. After an eligible failure, it can retry the same model through another deployment.

The optional perplexity/auto feature makes a different decision: it selects a model for each request and is described as a preview. Custom routing can pair a default model with one to five routes. Because the model can change between requests, its applicable token rate can change too.

Routing functionWhen the decision happensWhat can changeScope
Provider failoverWhen Orchestrator selects a deployment or retries after an eligible failureThe deployment serving the requestKeeps the already selected model
Optional model routingFor each requestThe model answeringperplexity/auto is a preview; custom routing supports one default model and one to five routes
If you need the same model across requests, distinguish deployment failover from perplexity/auto: the first changes where a model runs, while the preview can change which model answers.

Supported request formats and conversation history

Orchestrator accepts OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats. Perplexity says existing OpenAI or Anthropic SDK integrations can be redirected by changing the API key and base URL, then using a full model ID from its catalog. The quickstart documentation also describes the API as stateless: to continue a conversation, the client must send its history with each request.

Billing follows the published token rate of the model that answers, regardless of which provider deployment serves it. With optional model routing, check which model handled each request when estimating usage across a conversation.

What Perplexity’s reported results show

Perplexity says requests for Claude and GPT through Orchestrator succeeded more often than requests to any individual provider it measured, with less day-to-day variation during provider incidents. The company’s comparison covered October 4 through November 2, 2026, and excluded client errors such as invalid requests and customer rate-limit violations. These are Perplexity’s reported measurements. The company’s account of the comparison does not make them an independent test.

In a separate adjusted Terminal-Bench 4.1 comparison, Perplexity reported that each of two routed Claude Code configurations solved 41 of 63 tasks, while Claude Fable 5.2 alone solved 38. The company estimated a higher cost per task for the routed configurations.

Configuration in Perplexity’s comparisonTasks solved out of 63Estimated cost per task
Claude Fable 5.2 alone38Lower than the routed configurations
Each of two routed Claude Code configurations41Higher than Claude Fable 5.2 alone

The result is a company-reported comparison of those configurations: the routed setups solved three more tasks than Claude Fable 5.2 alone, alongside higher estimated cost per task.