AIGenerated
EN · IT · ES

Perplexity lets you send AI requests to 85+ models through one API with automatic backup switching

Companies can route every chat request to whichever model works best right now — switching automatically if one slows down — while paying just one monthly bill.

Apps and productsPerplexity lets you send AI requests to 85+ models through one API with automatic backup switchingHerald of AI

What you could do with it

  • A business building an AI app could send every request through a single API to 85+ models across dozens of providers, replacing months of juggling accounts, keys, contracts, and invoices with just one monthly bill 1.
  • A developer could point their existing OpenAI or Anthropic SDK at one base URL and start choosing from frontier and open-weight models immediately, only changing the API key and URL 12.
  • A team could monitor and control AI spending through a dashboard that breaks usage down by model, project, API key, and date, with optional spend caps and model allowlists 1.

One connection to rule them all

Until now, teams building applications with AI have needed individual accounts and API keys for OpenAI, Anthropic, Google and others. As new models arrive every week, those numbers only grow 1.

Anthropic and OpenAI alone released more than a dozen new models this year, alongside over a hundred open-weight models 1. Managing contracts and security reviews for each one quickly becomes impossible at scale .

The new Orchestrator API solves that by acting as a middle layer. Behind a single base URL it handles everything: choosing which model serves the request, sending it to the right provider, monitoring health and capacity, and switching providers automatically if one goes down 2. Requests support both the OpenAI Chat Completions and Responses formats plus the Anthropic Messages format, so most existing codebases need nothing more than a different API key and URL 1. The Anthropic SDK points to https://api.perplexity.ai/orchestrator without the /v1 suffix, while the OpenAI version includes it .

At a glance

One gateway, 85+ models behind it

  • OpenAIGPT models via openai.com
  • AnthropicClaude models directly
  • AWS BedrockClaude & GPT through Amazon
Orchestrator APISingle base URL · one monthly bill · auto-switches providers
  • AzureGPT up to 27 % faster than openai.com
  • Dozens of other providers85+ models total including open-weight options
  • Fallback routingIf one provider stalls, request retries on another deployment of the same model

From Perplexity announcement; Oct–Nov 2026 production traffic showed fewer bad days during provider incidents, though exact success-rate percentages were not pu

Failover built in, not DIY

If a request fails on one provider — because of rate limits, outages or slowdowns — Orchestrator retries it on another deployment of the same model 2. An invalid request still returns a clear HTTP 400 error and is never billed; only requests that actually start producing output get charged . If every backup also fails, the API returns a 429 status with a Retry-After header .

Amazon Web Services, Azure, and direct-provider endpoints were all measured against each other. During production traffic running from October 4 to November 2, Claude and GPT requests served through Perplexity achieved higher success rates than requests sent directly to any single provider tested 2. Daily success-rate variation during provider incidents was noticeably smaller too . The exact percentages were not published . Speed also differed: Azure's median output speed for GPT-5.6 Sol ran 27 percent faster than OpenAI's own endpoint, while Claude Opus 5.5 alternated evenly between Anthropic-direct and AWS Bedrock over 30 days . Full speed-aware routing remains "in progress" according to the announcement .

What developers can control now

In its current preview state the API offers two main tools. Routes let teams describe in plain language which model should handle each request type — choosing from Lower cost, Balanced or Higher quality profiles or building a custom rule set — with a default model handling anything unlabelled 1. Auto effort adjusts how much reasoning a model spends on its next turn according to policies the developer defines, so an agent can invest extra thinking after encountering a tool error without swapping models entirely .

Billing is unified under one vendor and one invoice, with no platform fee layered on top of usage credits 1. The API Console tracks spending broken down by model, project, key and date, supports per-project API keys, organisation-level allowlists, spend caps and capacity requests . Teams needing volume pricing can contact api@perplexity.ai directly .

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

How good is it

2 Perplexity reports that between October 4 and November 2, 2026, Claude and GPT requests routed via Orchestrator achieved greater success rates compared with those sent straight to individual providers that were evaluated. Daily success-rate variation during provider incidents was noticeably smaller, meaning fewer bad days when one provider had problems . The company did not publish exact percentage figures . Azure's median output speed for GPT-5.6 Sol ran 27 percent faster than OpenAI's own endpoint, while Claude Opus 5.5 split evenly between Anthropic-direct and AWS Bedrock over 30 days .

Key facts

  • One API connects to 85+ models across OpenAI, Anthropic, Google, xAI and open-weight options like DeepSeek and Kimi
  • Automatic failover retries failed requests on alternative providers without billing for unfinished attempts
  • Claude and GPT request success rates exceeded any single direct provider during Oct–Nov 2026 measurements
  • No platform fee is added on top of model usage credits; billing consolidates into one invoice

Sources

  1. Orchestrator: One API for frontier and open-weight models
  2. Perplexity Launches Orchestrator API for Failover and Model Routing — marktechpost.com

If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work

Keep reading