AI News
AI News AgentToolingPerplexity3 min read

Perplexity launches Orchestrator to use AI models

Perplexity has launched Orchestrator, an API that brings models from OpenAI, Anthropic, Google, xAI, and open-weight providers together under one connection and bill. The platform also manages routing, provider failures, and spending controls, with automatic selection features still in preview.

Perplexity has launched Orchestrator, an API that lets companies use models from OpenAI, Anthropic, Google, xAI, and several open-weight model providers through a single connection.

The idea is simple: not every task needs the most powerful model. An application can send a routine query to a cheaper model and save advanced models for complex problems. Orchestrator manages that decision and can also switch providers if one fails or runs out of capacity.

One connection for many models

Companies that integrate AI often end up with multiple accounts, access keys, contracts, and invoices. The problem grows whenever a new model appears: they have to decide whether to add it, renegotiate terms, and modify their infrastructure.

Orchestrator aims to bring all of that under a single provider and one bill. Its catalog includes models from OpenAI, Anthropic, Google, and xAI, as well as open-weight models such as GLM, Kimi, and DeepSeek. These models can run on Perplexity's infrastructure or through inference partners.

To get started, teams can continue using OpenAI and Anthropic's standard formats and SDKs. In many cases, they only need to change the API key, update the base URL, and select an available model.

Perplexity manages traffic and failures

The API is more than a model showcase. Perplexity says its production layer distributes hundreds of millions of requests across more than 85 models and more than a dozen providers.

Orchestrator uses that infrastructure to monitor each provider's health, capacity, and latency. If a request fails, it can be retried through another implementation of the same model without the team having to program all the recovery logic itself.

This matters especially for applications that process millions of requests a day. The difference between a 99% success rate and a 99.99% success rate can mean thousands of additional responses arriving correctly.

According to Perplexity, during one month of production traffic, Claude and GPT requests handled through its platform achieved higher success rates than those of any individual provider the company measured. They also showed less day-to-day variation during incidents.

Spending controls from a single dashboard

Orchestrator does not add a platform fee on top of credits, according to the company. It also includes tools to track and limit usage:

  • Spending reports by model, project, API key, and date.
  • Separate keys for each team or project.
  • Lists of models allowed for the organization.
  • User restrictions, spending limits, and capacity requests.

The practical benefit is that a company can change its model mix without taking on a separate commitment with each lab. For example, it can use a low-cost model to classify thousands of documents and a more capable one to draft the final reports.

Two features still in preview

Orchestrator is also testing two tools to automate model selection and reasoning effort.

Routes lets teams describe in natural language what types of requests each model should handle. They can choose a lower-cost, balanced, or higher-quality profile, or create their own policy. If no rule is defined, the default model is used.

Auto effort adjusts how much reasoning a compatible model uses on its next turn, based on a policy defined by the developer. This allows an agent to devote more effort to recovering from a tool error without having to switch models.

For you, the change will be most noticeable in the applications you use, not necessarily through a new interface. If Orchestrator achieves its goal, AI services will be able to switch models or providers behind the scenes to control costs and keep running when capacity problems arise. The important question will be whether that flexibility maintains consistent quality and whether the automated rules actually choose the right model for each task.