DIRTY STATIC / DISPATCH Tools19 NOV 2026 • 11 MIN READ
A concrete switchboard in a brutalist hall with one glowing caution yellow key on its front and cables running to a wall of concrete model blocks, hazard stripes on the floor
ToolsINTELLIGENCE FOR THE MACHINE AGE

Perplexity Orchestrator: One API, Many Models

Perplexity Orchestrator puts more than 45 models from OpenAI, Anthropic, Google and xAI, plus open weight models Perplexity hosts itself, behind one API key and one bill. This profile covers how its provider routing and its automatic routing work, what the catalog costs, which usage tiers gate which models, and where the limits sit.

Start reading ↓

A model outage used to be your outage. If the one provider behind your app slowed down, hit a rate limit or went dark, your users felt it in your product, and your only real option was to wait or to rebuild the integration under pressure.

Perplexity Orchestrator is built for that exact moment. It is a routing layer, meaning a service that sits between your app and the companies that serve models, and decides where each request should run. You name the model you want. Orchestrator picks a healthy place to run it, and if that place fails, it tries another place that serves the same model.

Perplexity Orchestrator

Official site

One endpoint for models from several labs, with routing and billing handled in one place. Primary use case: apps that call more than one model, or that cannot afford to fall over when a single provider has a bad day.

Key features: provider routing based on availability, capacity and latency; retry on another deployment of the same model after an error, a rate limit or a timeout; an optional automatic routing mode, in preview, that picks a model for each request from plain language task descriptions; and spend views by model, project and API key in the API Console.

Platforms
Web API, through the OpenAI SDK or the Anthropic SDK
Pricing
Per token at the served model published rate, no per request fee, from prepaid Perplexity API credits

What it is

Orchestrator is not a model. It is the serving layer Perplexity says it already runs for its own products, now opened to outside developers. Perplexity says that layer handles hundreds of millions of requests each day across more than 85 models and more than a dozen providers. The public catalog at introduction is smaller, at more than 45 models, and it mixes two kinds of supply.

The first kind is proprietary models from other providers, from OpenAI, Anthropic, Google and xAI. The second kind is open weight models, meaning models whose weights are public, from labs such as Z.AI, Moonshot AI and DeepSeek, hosted on Perplexity inference infrastructure. One key and one credit balance cover both kinds.

That mix is the interesting part. Teams often start with one frontier model, then add a cheaper open weight model for routine work, then add a second frontier model for one task the first one handles badly. Each addition used to mean a new account, a new key, a new bill and a new failure mode. Orchestrator turns that stack into one integration, which is a much easier thing to reason about on a tired afternoon.

It also sits in a wider pattern on this site. Kilo Desktop made the same bet inside a coding app, that the picker matters more than any single model. Orchestrator makes that bet at the API level, where the app itself is yours.

Two kinds of routing

The name hides two separate decisions, and keeping them apart makes the product much clearer.

Provider routing decides where a named model runs. You ask for a model by its full ID, such as openai/gpt-6-luna or perplexity/glm-5.3. Orchestrator distributes the request among deployments of that model using observed health and capacity. You do not choose the deployment, and the model does not change.

Automatic routing decides which model runs at all. It is optional and in preview. You set the model to perplexity/auto and supply a profile that pairs plain language task descriptions with models, plus a default model for everything else. A decision model reads the conversation, matches it to a route description without seeing the model names, and sends the request to the model assigned to that route. If nothing matches, or the decision model fails or times out, the default model answers.

Perplexity gives a support desk as its clearest example. Routine questions go to a fast and cheap default model. A billing dispute goes to a careful model. A broken integration goes to a strong troubleshooting model. As a conversation changes task, the model can change with it, turn by turn, while the full history travels with each request.

A concrete control desk with a glowing yellow dial and pipes branching to three concrete chambers, each holding a glowing yellow cube of a different size
One desk, several chambers. Provider routing picks the chamber for a named model, while automatic routing picks the model itself from the task you describe. Illustration: Dirty Static

Failover is the headline feature

Provider routing earns its keep when something breaks. If a provider returns an error, hits a rate limit or times out, Orchestrator retries on another deployment of that model, with limits on time to first token and total duration for each attempt. A request for Claude that fails on Anthropic can still succeed on AWS Bedrock, to use the example Perplexity itself gives, and it stays a Claude request the whole time.

There are edges, and Perplexity states them plainly in its documentation and coverage. Invalid requests and prompts that go past a model context window are client errors, not provider failures, and they do not get a retry on another deployment. If every deployment fails, the API returns HTTP 429 with a Retry-After header, and requests that fail before producing output are not billed.

The reliability claim behind all this is a company claim, and it should be read as one. Perplexity says that in its 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 it measured, and that daily success rates varied less during provider incidents. That is plausible, because more deployments give a request more places to land. It is also Perplexity measuring its own service, so treat the direction as the finding and the size of the gap as unproven until outside measurement exists.

Speed gets a similar honest note. Perplexity says the fastest provider for a model changes over time, giving Claude Opus 5.5 as an example where Anthropic led on 15 of 30 days and AWS Bedrock led on the other 15. Work on routing that adapts to live speed is still in development, by the company own account, so today the promise is resilience first and speed second.

A broken concrete bridge beside an intact one in a brutalist corridor, with a glowing yellow signal light guiding small yellow cubes across the intact bridge
Failover in one picture. When one route breaks, the work moves to another route that serves the same model, instead of stopping with the first failure. Illustration: Dirty Static

Catalog, tiers and platforms

Facts checked November 19, 2026 against the Perplexity Orchestrator blog post and the Orchestrator documentation.

Specifications

Public introduction
November 17, 2026
Catalog
More than 45 models at introduction, from OpenAI, Anthropic, Google and xAI, plus open weight models hosted by Perplexity
Formats
OpenAI Chat Completions, OpenAI Responses and Anthropic Messages
OpenAI base URL
https://api.perplexity.ai/orchestrator/v1
Anthropic base URL
https://api.perplexity.ai/orchestrator
State
Stateless, send the full conversation on every request
Billing
Per token at the served model rate, no per request fee, from the same prepaid credits as the rest of a Perplexity API project

Moving an existing integration is deliberately dull, which is a compliment. Change the base URL, swap in a Perplexity API key, and name a model by its full catalog ID. Every model supports all three request formats, though individual features depend on the model itself. A call to GET /orchestrator/v1/models returns the models your own project can call, with their base rates.

Access is tiered, and the tiers matter more than the headline catalog suggests. Google, xAI and open weight models work at every usage tier. OpenAI models and most Anthropic models require Tier 1, which starts at 50 dollars in cumulative credit purchases. The top Anthropic models, Claude Opus 5, Claude Opus 5.5 and the Claude Fable family, require Tier 2, which starts at 250 dollars. A request for a model above your tier returns HTTP 402 naming your current tier and the model minimum tier.

Pricing

There is no routing fee and no per request fee. You pay per token at the published rate of the model that answered, no matter which provider served it, so a failover does not change the price of that request. Usage draws from the same prepaid credits as the rest of your Perplexity API project, and the API Console breaks spend down by model and by API key.

Sample rates from the official catalog, in US dollars per one million tokens, checked November 19, 2026:

ModelMinimum tierInputOutput
perplexity/glm-5.3-flashEvery tier$0.15$0.50
anthropic/claude-sonnet-5-5Tier 1$2.00$10.00
anthropic/claude-opus-5-5Tier 2$4.00$20.00
anthropic/claude-fable-5-2Tier 2$8.00$40.00

The spread in that small table is the whole argument for routing. Routine work can run on a model that costs cents. Hard work can climb to a model that costs forty dollars per million output tokens. With one integration, that choice is a line in a routing profile. Without one, it is often a second vendor relationship that nobody got around to building, so everything runs on the expensive model by default.

Cache reads bill at a lower cache read rate, reasoning tokens bill at the output rate, and the documentation sets out cache write pricing and the flex and priority service tiers for OpenAI models. Those details reward a close read of the pricing page before a large rollout, because they move the real bill more than the headline input and output rates suggest.

What Works

  • One key, one base URL and one bill for models from four frontier labs plus open weight models Perplexity hosts
  • Failover retries a failed request on another deployment of the same model, so one provider outage does not have to become your outage
  • Existing OpenAI and Anthropic SDK code moves by changing the base URL, the key and the model ID
  • Automatic routing can match each turn of a conversation to a different model from plain language task descriptions

What Does Not

  • The strongest Anthropic and OpenAI models sit behind usage tiers that start at 50 and 250 dollars in cumulative credit purchases
  • Automatic routing is in preview, and routing that adapts to live provider speed is still in development
  • The reliability figures are Perplexity measurements of its own service, covering October 4 to November 2, 2026, with no outside measurement in the sources
  • Stateless requests mean long conversations are resent in full every turn, and server state fields return errors

Which should you choose

Choose Orchestrator when you already call, or expect to call, more than one model. The value compounds with each model you add, because each one arrives without a new account or a new bill. It compounds again if uptime matters, because failover across deployments is work very few small teams would build well on their own.

Choose Orchestrator when cost control is a routing problem. If most requests are routine and a few are genuinely hard, a profile that sends routine work to a cheap open weight model and hard work to a frontier model can cut spend without cutting quality where it counts. Perplexity own research products, covered here in Perplexity Teammates, show how much work the company now expects agents to hold for long stretches, and long running work is exactly where per request model choice adds up.

Stay direct when you live on one provider and you are happy there. One integration you control, with a direct relationship and direct support, is a fine place to be. Orchestrator adds a layer between you and the model, and a layer only earns its place when it removes more work, risk or cost than it adds.

Final thoughts

The clever move here is not the catalog. Plenty of services can list models. The clever move is separating the model you asked for from the place that happens to serve it, then treating that separation as the default rather than as an advanced setting. Most outages, rate limits and latency spikes live in that gap, and Orchestrator is built to absorb them quietly while the bill stays tied to the model, not to the accident of which provider was healthy.

Its limits are just as clear. Tiers gate the models people will ask about first, automatic routing is a preview, and the reliability story rests on the maker own numbers for now. None of that undermines the shape of the product. One API, many models, with failover when a provider fails, is a shape a lot of teams were already trying to build by hand. Perplexity is now selling them the finished version.

Sources

■

END OF TRANSMISSION

Stay curious. Question everything.

CONTINUE THE DESCENT

MORE FROM DIRTY STATIC ↘