Perplexity's Orchestrator API Keeps Your Model, Switches Providers
Perplexity's Orchestrator API, launched November 17, 2026, retries failed requests on another deployment of the same model and beat every individual provider on success rates.
By Grace Kim
4 min read

What's News
- Perplexity launched Orchestrator on November 17, 2026 with same-model provider failover and consolidated billing.
- From October 4 to November 2, 2026, Claude and GPT requests via Orchestrator succeeded more often than on any individual provider measured.
- Azure's median GPT-5.6 Sol output speed was 27% higher than OpenAI's over the 30-day window.
- Tier 0 GLM-5.3 Flash costs $0.15 per million input tokens and $0.50 per million output tokens.
- Automatic model routing via perplexity/auto is in preview and can change models, capability and cost between turns.
Perplexity's new API routed Claude and GPT requests with higher success rates than any individual provider it measured, according to data the company published on November 17, 2026. That claim anchors Orchestrator, the new API Perplexity launched the same day for accessing proprietary and open-weight models on its own inference infrastructure.
Orchestrator bundles three things: consolidated billing, same-model provider failover, and an optional automatic model-routing feature currently in preview. The pitch is simple — resilience without changing the model you picked or the per-token rate you pay.
How does same-model failover work?
Perplexity's routing documentation splits routing into two decisions. Provider routing comes after model selection and distributes traffic across deployments using observed error rates, capacity and latency. Customers don't choose the deployment.
When a provider error, rate limit or timeout hits, Orchestrator retries the request on another deployment of the same model. A Claude request that fails at Anthropic can move to AWS Bedrock and still be a Claude request. The numbers back the design: between October 4, 2026 and November 2, 2026, Claude and GPT requests through Orchestrator succeeded more often than requests to any single provider Perplexity measured, with smaller daily swings during provider incidents.
The speed spread across providers is wide. Over the same 30 days, Anthropic and AWS Bedrock each led Claude Opus 5.5 output speed on 15 days. Azure's median GPT-5.6 Sol output speed ran 27% higher than OpenAI's. Adaptive routing that exploits those speed differences is still in development.
The takeaway: one API key buys you failover across providers that a direct provider connection can't match.
What are the limits?
Failover has edges. Invalid requests and prompts exceeding a model's context window return HTTP 400 with no retry on another deployment. If every deployment is exhausted, Orchestrator returns HTTP 429 with a Retry-After header.
Billing rules follow the failure point. Requests that fail before producing any output aren't billed. If a stream dies after output begins, the customer pays for the tokens already delivered.
Perplexity counted failures conservatively in its measurements: a failure meant every available deployment was exhausted or the serving layer returned an error. Client errors such as invalid requests and client rate-limit violations were excluded.
Which models can you call?
The catalog spans model families from Anthropic, OpenAI, Google, xAI, Z.AI, DeepSeek and Moonshot AI. Access tiers differ:
- Google, xAI and open-weight models: available at every usage tier
- OpenAI and Anthropic models: Tier 1 or higher
- Claude Opus 5, Opus 5.5 and Fable models: Tier 2 or higher
Integration is deliberately frictionless. Orchestrator accepts OpenAI Chat Completions, OpenAI Responses and Anthropic Messages requests. Developers keep their existing OpenAI or Anthropic SDK and swap in a Perplexity API key, the Orchestrator base URL and a full model ID from the catalog.
One distinction matters for architects: the quickstart describes Orchestrator as stateless and built for direct model access with customers' own prompts and tools. That separates it from Perplexity's Agent API, which returns web-grounded answers with built-in citations.
What does automatic routing change?
The preview feature targets model choice, not deployment choice. Setting the model to perplexity/auto invokes a decision model that sees conversation text and plain-language route descriptions — but not route names or model IDs. A custom routing object can define one default model plus one to five task routes. The default model takes over when no route fits, when the decision model fails or times out, or when the conversation includes images, audio, video or files.
Routing runs independently per request, so turns within a single conversation can land on different models as the task changes. Two caveats follow. Every request must carry the full conversation, and each model maintains a separate prompt cache. Stable session identifiers or prompt-cache keys can keep related requests on one deployment, but that affinity is best effort and doesn't pin model selection across turns. Response headers describe the routing outcome; the response's model field names the model that actually answered.
Pricing follows the model that answered, at its published per-token rate regardless of which deployment served it. Usage draws from the same prepaid credits as the rest of a Perplexity API project. The cheapest catalog entry, Tier 0 GLM-5.3 Flash, lists at $0.15 per million input tokens and $0.50 per million output tokens, with cache reads at $0.03 per million tokens. Reasoning tokens bill at the applicable output rate.
The reliability story is concrete today; the automation story isn't. Until adaptive, speed-aware provider selection ships, customers buying Orchestrator for uptime get a proven product — those buying it for hands-off model switching get a preview whose capability and cost can shift between turns.
Source: unite.ai
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,301 articles