Beating AI News Flash: Perplexity has launched the Orchestrator API, integrating over 45 models including GPT, Claude, Gemini, Grok, as well as GLM, Kimi, DeepSeek, and more into a single interface. Developers only need one set of API keys to call models from different companies, with unified billing.
It provides two layers of automatic orchestration. The first layer addresses service stability: for example, if Anthropic's Claude API returns errors, rate limits, or times out, the system attempts to switch to the same Claude model provided by AWS Bedrock, without requiring developers to handle it manually.
The second layer automatically selects models based on the task. Developers can specify rules in natural language, for example, routing simple code modifications to the cheaper GLM-5.3 Flash, complex troubleshooting to Claude Opus 5.5, and cross-module design to GPT-6.1 Sol. The system reads the current conversation, and a dedicated decision model determines which model to use. Even if the model is switched mid-way, it can continue reading the full prior conversation. If it later switches back to the original model, it can also attempt to reuse its previously left cache to reduce redundant computation.
Perplexity also tested a special use case: when Claude refuses a normal task due to safety restrictions, switching to GPT to continue processing. The team set switching rules for cybersecurity, biology, and post-refusal requests. In the 63 coding problems of Terminal-Bench 4.1, Claude Fable 5.2 completed 38 on its own, and with routing added, completed 41, but the estimated cost per task also increased.
