It was the kind of night every developer dreads. Your app’s AI feature stops answering, the error log fills with rate-limit messages, and somewhere on a status page you have never bookmarked, your provider admits something is wrong. Perplexity Orchestrator is built for exactly that night.
Perplexity has opened up the model-routing layer it already uses in production. The new Orchestrator API puts more than 45 AI models from OpenAI, Anthropic, Google and xAI, plus open-weight models Perplexity hosts itself, behind a single endpoint. One key, one bill, and if a provider goes down, your request quietly moves to another deployment of the same model.
The launch was announced on November 17, 2026, in a Perplexity blog post titled “Orchestrator: one API for reliable inference”. A private preview had already been running since November 9. According to Perplexity, the service is the first step in a wider API platform built around intelligent model orchestration.
One API, 45+ Models: How Perplexity Orchestrator Works
The core idea is simple. Instead of wiring your app to four or five different providers, each with its own SDK, key and billing page, you point your existing OpenAI or Anthropic SDK at one base URL and use your Perplexity API key. The API accepts OpenAI Chat Completions, OpenAI Responses and Anthropic Messages request formats, so most teams only need to change the key and the URL.
Behind that single endpoint sits a catalogue of more than 45 models. The proprietary side covers the big names: OpenAI, Anthropic, Google and xAI. The open-weight side is hosted by Perplexity itself and includes Kimi K3 from Moonshot AI, GLM-5.3 and GLM-5.3 Flash from Z.AI, and DeepSeek V4 Pro from DeepSeek. New models are added as they ship, without any code change on the customer’s side.
Perplexity says the infrastructure behind Orchestrator already handles hundreds of millions of requests a day across more than 85 models and over a dozen providers. That is the same routing layer its own products have been running on, now sold to outside developers.
When a Provider Fails, Perplexity Orchestrator Switches Automatically
The feature that matters most is failover, and it works in a specific way. Orchestrator keeps the model fixed and swaps the deployment. If a Claude request fails on Anthropic’s own servers, it can be served by AWS Bedrock instead, as long as the same model is available there.
Routing decisions come from live observation. Orchestrator tracks the health, capacity and latency of every provider, then distributes traffic across the best-performing deployments. A provider error, a rate limit or a timeout triggers a retry on another deployment of the same model.
Each attempt has its own limits for time to first token and total duration. Exceed either limit and that attempt counts as a failure, so a slow provider does not hold your user hostage. If every deployment fails, the API returns an HTTP 429 response with a Retry-After header.
There is also a fairness detail worth noting. Requests that fail before producing any output are not billed. You pay for answers, not for attempts that went nowhere.
The Reliability Claim: Higher Success Rates Than Any Single Provider
Perplexity backs the pitch with a month of its own production traffic. In measurements from October 4 to November 2, 2026, Claude and GPT requests sent through Orchestrator succeeded more often than requests sent to any individual provider Perplexity measured.
The same data showed less daily variation during provider incidents. In other words, the bad days got smoother. When one provider stumbled, traffic moved elsewhere and the success rate barely moved.
That claim deserves a careful reading. It comes from Perplexity’s own measurements, not an independent audit, and it covers a single 30-day window. But it points at a real problem: the week of November 12 to 17 saw outages hit Microsoft 365, Apple, half a dozen US banks, United Airlines and even a classified NSA network.
Orchestrator’s answer to that fragility is redundancy by default. Instead of building your own failover logic, retry queues and provider health checks, you rent them from a company whose whole business is routing requests.
One Bill, No Platform Fee
The billing model is deliberately plain. Each request is charged at the answering model’s published per-token rate, regardless of which provider actually served it. Provider failover does not change what you pay.
Usage draws from the same prepaid credits as the rest of a Perplexity API project, and spending can be viewed by model and filtered by API key in the API console. Credits can also be bought through AWS Marketplace.
Most striking is what Perplexity does not charge. There is no platform fee on top of credits. That is a direct contrast with OpenRouter, the best-known model gateway, which charges a 5.5 percent standard fee on credit purchases. For a team already spending heavily on Claude and GPT usage, the difference adds up fast.
Access is tiered by cumulative credit purchases. Tier 0, at zero dollars, covers Google, xAI and open-weight models. Tier 1, at 50 dollars, unlocks OpenAI and most Anthropic models. Tier 2, at 250 dollars, opens Claude Opus 5, Opus 5.5 and the Fable models.
Routes: Describe the Job in Plain Language
Orchestrator also offers a second routing layer, currently in preview. Automatic routing chooses the model itself, not just the provider. Developers set the model to perplexity/auto and supply a routing object with a default model plus up to five named routes.
Each route pairs a plain-language description of a task with a model. A decision model reads the conversation and the route descriptions, never the route names or model IDs, and picks where each request goes. Simple, high-volume work can run on a cheaper model while demanding work goes to a stronger one.
The trade-off is control. Orchestrator is stateless, so every request must include the full conversation, and the routing decision is made for you, which means trusting Perplexity’s decision model to read your traffic correctly.
Auto Effort: Reasoning That Adjusts Itself
Alongside Routes, a second preview feature called Auto Effort adjusts how much reasoning a supported model uses on its next turn. The adjustment follows a policy the developer defines, so a team can spend more thinking on hard questions and less on easy ones.
It is a small feature with a large implication. Reasoning models are priced by the token, and most of their cost sits in the thinking nobody sees. A policy that trims reasoning where it is not needed can cut bills without touching the final answer.
The Catch: What Perplexity Orchestrator Does Not Let You Do
Orchestrator is not a free-for-all. It does not support bring-your-own-key, so you cannot plug in your own Anthropic or OpenAI credentials. You cannot manually pick a deployment either. If you want a request to run on a specific provider for compliance or latency reasons, that choice is not yours to make.
Access is also still gated. The service is in private preview, and developers who want in during that phase need to email Perplexity’s API team. The tiered model access means the most capable models require a spending history first.
Then there is the question of dependence. Orchestrator removes the risk of any single provider failing, but it concentrates that risk in Perplexity itself. If Perplexity’s routing layer goes down, every model behind it goes down with it. That is a trade most teams will accept, but it is a trade worth naming.
Why This Launch Matters Now
Perplexity is arriving at a crowded moment. Model gateways are not new, and OpenRouter has spent years building the category. What is new is who is doing the routing: a company with its own search product, its own agent platform and, as of this month, its own Teammates agents for ongoing work.
The timing also fits the industry’s mood. After a year in which AI models multiplied faster than anyone’s ability to evaluate them, the bottleneck has moved from choosing a model to keeping it running. Developers no longer ask which model is best. They ask which one will answer at 3 a.m.
Orchestrator’s answer is that you should not have to care. One endpoint, one bill, and a router that moves your request before you notice it failed. If Perplexity can keep that promise through the next big outage, the gateway business may turn out to be as valuable as the models behind it.



