Perplexity has launched the Orchestrator API, a single access point to proprietary and open-weight models, the latter with downloadable weights. For developers, it means connecting several providers without managing several integrations or invoices.
What it is and how it connects
Orchestrator is a hosted inference gateway: it answers at one base address and routes each request to the chosen model. The catalogue includes models from OpenAI, Anthropic, Google and xAI, alongside open families such as GLM, Kimi and DeepSeek, run by Perplexity or by specialist inference partners. It accepts requests in the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats, so in most cases only the key and address change. Perplexity runs the same layer for its own applications, which cover more than 85 models from over a dozen providers.

Image: Perplexity — https://docs.perplexity.ai/docs/getting-started/overview
What happens when a provider fails
Failover keeps the model fixed and changes the infrastructure serving it. A Claude request that fails at Anthropic can be served by AWS Bedrock and remains a Claude request. The switch triggers on provider errors, rate limits or timeouts. Malformed or oversized requests get a 400 error and are not retried; if every infrastructure fails, the reply is a 429 indicating when to retry. Requests that fail before producing output are not billed, and developers cannot choose the infrastructure themselves. The time-to-first-token and total-duration limits are not published.
The reliability figures
Perplexity says that between 4 October and 2 November it measured live traffic: Claude and GPT requests succeeded more often than with any single provider it tracked, and daily success rates varied less during outages. The post gives no precise percentages, no methodology and no provider names, and no third party has verified it. Speed also varies by provider: for Claude Opus 5.5, Anthropic was fastest on 15 of 30 days and AWS Bedrock on the other 15; for GPT-5.6 Sol, Azure's median output speed was 27% higher than OpenAI's. Speed-aware routing is still in development.
Automatic routing, in preview
The most ambitious mode is perplexity/auto, still in preview. The developer sets a default model and a list of routes, each with a name, a description and a model. A decision model reads the conversation and the route descriptions, but never sees route names or model IDs. If no route matches, or the decision model fails or times out, the default answers, including for media inputs. Response headers report which route was chosen and with what confidence. The default policy sends traffic to GPT-6.1 Sol and simple requests to GPT-6 Luna; low-cost, balanced and high-quality presets are also available.
Can routing get around model refusals?
Perplexity tested the approach on 63 security and life-sciences tasks run in Claude Code. Claude Fable 5.2 alone completed 38; adding routes to GPT-6.1 Sol or GPT-6 Astra raised that to 41. The company argues that Fable's safeguards block legitimate work, but the gain is three tasks at a higher estimated cost per task, and Perplexity itself calls it an adjusted comparison: results vary by workload.
Pricing and access
Each request is billed at the answering model's published per-token rate, and switching provider does not change it. Perplexity says the platform adds no fee, while some rival services take a percentage on top-ups. Usage draws on prepaid API credits, also purchasable through AWS Marketplace. Access depends on the tier: the first covers Google, xAI and open-weight models; the second, from $50, adds OpenAI and most Anthropic models; the third, from $250, unlocks Claude Opus 5, Opus 5.5 and the Fable models.
Against OpenRouter and Vercel
OpenRouter, which lists over 400 models from more than 70 providers, and Vercel's gateway offer a similar service. OpenRouter charges 5.5% on credit purchases on its standard plan and 8% on its business plan; Vercel charges no platform fee and no token markup. Perplexity does not state an exact model count.
The launch fits a trend we have already followed: Cloudflare has put a router that picks the model for each request into beta, and Architect has brought auctions between providers to single requests. US companies are also diversifying their models to contain costs, and AnonRouter promises to route without reading prompts.
For teams adopting these services, the question is whether the claimed reliability holds on their own traffic, and whether automatic routing keeps its promises on quality, since the published evidence comes from the provider itself.

No comments yet. Be the first!