Skip to content
NLEN
Illustration: OpenAI pipeline found 948 Linux root bugs in one run: what autonomous bug hunting means

Perplexity Orchestrator API: 45 models behind one key

By Ivo Donker · — text: Claude (AI)

What is Perplexity's Orchestrator API? It is an API that lets you call models from Anthropic, OpenAI, Google, xAI and others with a single Perplexity key and a single bill. The model catalogue lists 45 models. You pay each model's per-token rate, with no per-request fee. If a call fails, Perplexity retries it on another deployment of the same model. Your own OpenAI or Anthropic keys are not accepted.

We read the FAQ and the models and pricing page in Perplexity's documentation. The launch was announced on X on November 17 in a post by Perplexity. The topic was also covered in the trending briefing of November 18.

What happened

Perplexity is opening its own model routing to developers. Anyone with a Perplexity API key can use the same key for frontier and open-weight models, with consolidated billing. The API speaks three formats: OpenAI Chat Completions (/chat/completions), OpenAI Responses (/responses) and Anthropic Messages (/messages). According to the documentation every catalogue model supports all three, although features vary by model.

With the model name perplexity/auto, a decision model picks which model answers each request, following a default policy or routes you define yourself. You are billed at the rate of the model that answered; perplexity/auto itself does not appear in the price tables.

Details

Details
ItemAccording to the documentation
Models45: Anthropic 15, OpenAI 18, Google 5, xAI 3, Z.AI 2, DeepSeek 1, Moonshot AI 1
PricingPer token at the model's rate; no per-request fee; reasoning tokens at the output rate
ExampleClaude Sonnet 5.5: 2 dollars input, 10 dollars output per million tokens
FailuresRetry on another deployment of the same model; if all fail, HTTP 429 with Retry-After
Bring your own keyNot supported
AccessGoogle, xAI and open-weight models from Tier 0; Anthropic and OpenAI from Tier 1; Opus 5, Opus 5.5 and Fable from Tier 2

Requests that fail before producing output are not billed, and neither are requests rejected by the rate limit. Invalid requests and requests that overflow the context window are not retried. Responses carry no cost field, so you cannot read the price of a call from the response itself. Organisations with zero data retention cannot call the Fable models. The new GPT-6.2 models were not yet in the list when we read it.

What this means

For developers who want to switch between models without running their own gateway, this is a ready-made route with one invoice and automatic failover. The trade-offs: you cannot bring existing provider contracts or credits, and the most expensive models require a higher usage tier. Because responses do not report cost, you have to track spending yourself or check the dashboard.

Until now Perplexity mainly built products on top of this routing, as described in our article on Perplexity Teammates and the article on its Decider routing model. To keep costs down in such a setup, read our article on context caching. What agents do with APIs like this is covered in our overview of agentic AI.

What we did not verify

The post mentions spend limits and allowed-model lists per key; we did not find those in the documentation we read. How perplexity/auto makes its choice, and whether that choice itself costs anything, is not described. We did not check the revenue claim by CEO Aravind Srinivas on X. We did not test the API ourselves.

Checked on November 18, 2026.

This article was written with AI (Claude) and checked against the sources below.

Sources