Perplexity has launched the Perplexity Orchestrator API, a single endpoint for frontier and open-weight models. It picks a serving provider by availability, capacity and latency, and bills everything through 1 account. Developers get fewer integrations and 1 invoice.
What is the Perplexity Orchestrator API?
The Perplexity Orchestrator API is a hosted inference gateway that serves many models behind 1 base URL. It covers frontier models from OpenAI, Anthropic, Google and xAI. It also serves open-weight models such as GLM, Kimi and DeepSeek, hosted by Perplexity or inference partners.
It accepts OpenAI Chat Completions, OpenAI Responses and Anthropic Messages requests. Most teams only change the API key and base URL. The OpenAI base URL is https://api.perplexity.ai/orchestrator/v1. Anthropic SDK users point at https://api.perplexity.ai/orchestrator, without /v1.
Perplexity runs the same layer internally, across 85+ models and more than a dozen providers.
How does provider failover work?
Failover keeps the model fixed and swaps the deployment. A Claude request that fails on Anthropic can be served by AWS Bedrock instead. The reliability docs set these rules:
- Triggers are provider errors, rate limits and timeouts.
- Each attempt has a time-to-first-token limit and a total-duration limit. The values are not published.
- Invalid or over-context requests return HTTP 400 and are not retried.
- If every deployment fails, the API returns HTTP 429 with
Retry-After. - Requests that fail before producing output are not billed.
You cannot pick a deployment yourself.
What do the reliability and speed numbers show?
Perplexity measured production traffic from October 4 to November 2, 2026. Claude and GPT requests through Perplexity succeeded more often than requests to any single provider measured. Daily success rates also varied less during provider incidents. The post gives no percentage success rates.
Speed varied by provider too. For Claude Opus 5.5, Anthropic was fastest on 15 of 30 days, AWS Bedrock on 15. Azure’s median output speed for GPT-5.6 Sol was 27% higher than OpenAI’s. Speed-aware routing is still in progress.
How does automatic routing with perplexity/auto work?
Routing is in preview. You set model to perplexity/auto and pass a routing object. It holds a default_model and 1 to 5 routes, each with a name, description and model.
A decision model reads the conversation and route descriptions. It never sees route names or model IDs. If nothing matches, or the decision model fails or times out, the default model answers. So do media inputs.
Response headers report the result: x-perplexity-autoroute-outcome, -route and -confidence. The default policy sends traffic to GPT-6.1 Sol, with simple requests going to GPT-6 Luna. Perplexity also lists built-in Lower cost, Balanced and Higher quality profiles, plus an Auto effort preview.
