Perplexity Orchestrator API: One Key for GPT and Open Models
Perplexity launched Orchestrator on November 17, 2026: one key for GPT, Claude, and open models, with failover. Higher tiers gate OpenAI and Anthropic.

Perplexity launched Orchestrator on November 17, 2026: one key for GPT, Claude, and open models, with failover. Higher tiers gate OpenAI and Anthropic.

On November 17, 2026, Perplexity published Introducing Orchestrator: One API for frontier and open-weight models ↗ and posted it on X ↗. The mechanics below are from the docs fetched November 18: the quickstart ↗, the models page ↗, Routing & Reliability ↗, and the FAQ ↗. aiholics ↗ (Alex Carter, November 17), Crypto Briefing ↗ (Diego Almada Lopez, November 17), and TokenPost ↗ (Simon Yoon, November 17) covered the launch. The picture above is an original graphic created by aitoolranked. We have not sent a request to this API.
The quickstart is one key you already have, one bill, and three request shapes: OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. A switch is the base URL plus a catalog id such as openai/gpt-6-luna. OpenAI-style calls use https://api.perplexity.ai/orchestrator/v1. The FAQ gives Anthropic SDKs https://api.perplexity.ai/orchestrator, because those clients append /v1/messages. Set the model to perplexity/auto and a decision model picks per request. The API is stateless. The FAQ says store: true, previous_response_id, and conversation on Responses each return HTTP 400.
Perplexity's announcement says the catalog covers OpenAI, Anthropic, Google, and xAI, plus open-weight models such as GLM, Kimi, and DeepSeek, served by Perplexity or by inference partners. The models page lists those families with a per-token rate and a minimum tier. The OpenAI table on that page, fetched November 18, has no openai/gpt-6.2-astra and no openai/gpt-6.2-sol. Perplexity's separate post the same day put GPT-6.2 in Search and Computer, which is not this API. Crypto Briefing and TokenPost date a private preview to November 9, 2026, and still describe access by email. The docs fetched November 18 tell an existing keyholder to change the base URL. The announcement sends large or committed workloads to api@perplexity.ai for volume pricing. The FAQ uses that address for models off the catalog.
The same day, Perplexity posted a Multiagent effort level ↗ in Computer and for teammates.
Google, xAI, and open-weight models start at Tier 0. OpenAI and Anthropic need Tier 1 or higher. Claude Opus 5, Opus 5.5, and the Fable models need Tier 2 or higher. A model above your tier returns HTTP 402. With zero data retention on, the FAQ says the Fable ids return HTTP 403. Calls use the Perplexity key and Perplexity credits, at the models-page rates, with no per-request fee. The announcement says there is no additional platform fee on top of credits. aiholics notes that the announcement does not publish per-model prices, rate limits, latency overhead, regional availability, or data-handling terms. The models page does publish rates. The FAQ does describe a requests-per-second limit in response headers, and zero data retention. This piece does not compare those rates with OpenRouter.
Once the model is chosen, the routing page says Perplexity picks a deployment from error rates, capacity, and latency, and retries a failure on another deployment of the same model when one exists. An invalid request, or a prompt past the context window, is HTTP 400 and is not retried. If every deployment fails, the result is HTTP 429 with Retry-After. A request that fails before any output is not billed.
perplexity/auto takes an optional routing object: a default model and one to five routes, each with a plain-language description. The decision model reads the description. With no object, the documented default sends simple requests to openai/gpt-6-luna and uses openai/gpt-6.1-sol otherwise, and both need Tier 1. Perplexity's announcement calls Routes and Auto effort preview features and says teams can start from built-in Lower cost, Balanced, or Higher quality profiles. The routing page documents "reasoning_effort": "auto", which picks low, medium, or high for that turn. The routing docs page fetched November 18 does not list those three profile names.
"Today, we’re launching our Orchestrator API, that picks the right model for your task: balancing across capability, capacity and latency" — Aravind Srinivas ↗ (@AravSrinivas), Perplexity's CEO, speaking for the company.
"Set perplexity/auto and a decision model will pick the model for each request." and "I had no idea this was happening lol" — Julian Brooks ↗ (@OutsourcedLogic)
"you can use different AI models for different jobs and get a single bill" — John Savage ↗ (@johnsavage_ai)
Reddit was not reachable for this story.
One invoice, a Messages-compatible client, and a retry that stays on the same model id are the reasons to try it. Failover, as the docs describe it, does not move you to a different lab when a deployment fails. Hugging Face's move onto OpenRouter is the other router change from November 16, and OpenRouter Auto Code is the coding router from November 8. Price a real workload against the models page, and time it, before you move production. Direct vendor APIs stay in that comparison. This piece does not rank Perplexity's rates against OpenRouter's.
The tier gate is the line under the slogan. Tier 0 reaches Google, xAI, and the open-weight ids. OpenAI and Anthropic start at Tier 1. Fable, Opus 5, and Opus 5.5 start at Tier 2, and the default auto policy needs Tier 1 because it names GPT-6.1 Sol and GPT-6 Luna. GPT-6.2 Astra and Sol are missing from the OpenAI table we fetched. A stateless Responses API also rejects the fields that assume the server stored the thread. Send the whole conversation on every turn.
Write your own route descriptions until the named profiles show up in the account. The docs support custom descriptions and automatic effort. The announcement's Lower cost, Balanced, and Higher quality profiles were not on the routing docs page we fetched. Decider returns probabilities and spans. Teammates is the hosted Max agent. The product page on this site is the Perplexity review.
Verdict: a sound trial if you want one bill and an Anthropic-format client. Leave production on the current provider until a week of your own traffic has a price and your tier can call every model you named.