Perplexity released an Orchestrator API on November 17 that routes requests across models and fails over automatically when one degrades. Reported at the same time: an annualised revenue run rate around $1 billion, roughly four times the figure at the start of the year.
The product
Orchestrator sits in front of multiple models and decides which one handles a given request. The headline feature is failover — when a model returns errors or degrades in latency, traffic shifts to another without the caller handling it.
That is unglamorous infrastructure, and it is exactly what production systems need. Every team building on a single provider eventually writes this layer badly, at 2am, during an outage. Buying it is rational.
Why failover is harder than it sounds
The obvious version — health check, switch on error — fails in the cases that matter. Model degradation is rarely a clean 500. It is a slow drift in latency, a rise in refusals, or answers that get shorter and less grounded. None of those trigger an error handler.
A useful orchestrator has to define quality signals per task: did the answer cite sources, did the JSON parse, did the response satisfy a schema. Then failover becomes a policy on those signals rather than on HTTP status codes.
Perplexity has not published its quality criteria, which is the first thing to ask about. A router that fails over on latency alone will happily serve worse answers faster.
The revenue number
$1 billion annualised, up roughly fourfold in a year, is a meaningful data point for a category that has been written off repeatedly. It says that answer engines can monetise, at least at consumer scale, and that the subscription model works better for this product than advertising did.
It also reframes the competitive question. Perplexity is no longer only competing with Google's answer box; it is selling infrastructure to developers, which puts it in a different market with different margins.
The strategic tension
Selling orchestration while running a consumer answer engine creates a conflict: Perplexity's own product is the best reference customer, and its routing decisions are therefore both a product and a sales pitch.
That is normal — AWS does the same — but it means the API's roadmap will be shaped by consumer economics. If the consumer product needs cheap models for the bulk of traffic, the API will be optimised for cost routing. Enterprise buyers who need quality routing should check which direction the defaults point.
What to test
Ask three things before adopting it. What signals trigger failover, and can you define your own? What happens to the prompt — is it rewritten per model, and does that change output format? And what is the added latency, because a router in the request path costs milliseconds on every call, not just the failures.
The honest read
Orchestration is becoming a layer rather than a feature, and the companies that own it will take a margin on every request without owning a model. That is a good business if the reliability claims hold, and a commodity if every provider ships the same thing — which, given how quickly the frontier labs are adding routing, is the more likely outcome within a year.