Models from several companies through one API and one bill
Perplexity on Nov. 17 (Nov. 18 Japan time) announced the "Orchestrator API," which gives access through one API to models from OpenAI, Anthropic, Google and xAI, as well as open-weight models such as GLM, Kimi and DeepSeek. Usage is consolidated into one bill, and Perplexity handles switching when a provider runs into trouble. The company said it expects teams to run simple, high-volume tasks on lower-cost models and reserve more powerful models for more demanding work.
The official @perplexity_ai account posted at 5:13 a.m. Nov. 18 Japan time that open-weight and proprietary models are available in one API with unified billing.
Introducing the Perplexity Orchestrator API. Orchestrator lets you choose the right model for the right task, and automatically selects the provider based on availability, capacity, and latency. All open-weight and proprietary models are in one API with unified billing.
According to Perplexity, labs such as Anthropic and OpenAI have released more than a dozen new models this year alone, and more than 100 open-weight models have also shipped. The company says that while teams have more options, they also have more providers, contracts and invoices to manage. Perplexity's own model layer routes hundreds of millions of requests across 85+ models and more than a dozen providers, and the company said it lowered its own serving costs by moving workloads to models suited to each task. The Orchestrator API opens that infrastructure to outside developers.
Existing SDKs work after changing the key and the base URL
The Orchestrator API supports three request formats: OpenAI's "Chat Completions" and "Responses," and Anthropic's "Messages." According to the documentation, every model in the pricing table can be called in all three formats, though available features vary by model. Integrations built with OpenAI or Anthropic SDKs can be moved by changing the base URL and the API key and specifying model IDs prefixed with the developer's name, such as openai/gpt-6-luna or perplexity/glm-5.3.
The service uses existing Perplexity API keys. The API is stateless, meaning it keeps no conversation state on the server, and request fields that depend on server-side state return an error. Perplexity's API lineup also includes the "Agent API," which answers with web search results and citations, and the "Decisions API," which returns probabilities for yes/no and similar questions. The documentation says the Orchestrator API is for direct model access with your own prompts and tools. In the Decisions API line, the company has also released an app that runs a small decision-only AI on Macs.
45 models from seven companies, no extra platform fee
The documentation's pricing table lists 45 models: 15 from Anthropic, 18 from OpenAI, 5 from Google, 3 from xAI, 2 from Z.AI (Zhipu), 1 from DeepSeek and 1 from Moonshot AI (ORIENS counted the table rows as of about 6:22 a.m. Nov. 18 Japan time). Anthropic's lineup runs from "Claude Fable 5.2" to "Claude Haiku 4.5," and OpenAI's from "GPT-6.1 Sol" to "GPT-5 nano." Google is represented only by five Flash models such as "Gemini 3.8 Flash." The open-weight "GLM-5.3," "DeepSeek V4 Pro" and "Kimi K3" carry the perplexity/ prefix. According to the blog, open-weight models are served by Perplexity or through inference partners.
Requests are billed at the published rate of the model that answered, regardless of which provider served them. The blog said, "There's no additional platform fee on top of credits." Perplexity cited as a benefit that teams can change their model mix without renegotiating contracts with each provider. It did not compare its prices with using each company's API directly, nor did it give actual savings.
Available models depend on usage tier
Which models can be used depends on the API usage tier. Google, xAI and open-weight models are available at every tier, OpenAI and Anthropic models require Tier 1 or higher, and Claude Opus 5, Opus 5.5 and Fable require Tier 2 or higher. Fable models are not available to organizations with zero data retention enabled.
Spending is managed in the "API Console." Spend is broken down by model, project, API key and date, and separate API keys can be issued per project or team. Administrators can set organization model allowlists, user restrictions, spend caps and capacity requests.
Teams with large or committed workloads can discuss volume pricing and capacity individually. Regional availability and the conditions for moving up a usage tier are not described in the materials ORIENS checked.
Automatic retries during provider outages; success rates measured by Perplexity
When the same model is served by several providers, the Orchestrator API distributes traffic based on each provider's health, capacity and latency. According to the documentation, when a provider returns an error, a rate limit or a timeout, the request is retried on another deployment of the same model. Invalid requests and prompts that exceed the model's context window are not retried. If every deployment fails, the API returns an error, and requests that fail before producing any output are not billed.
Perplexity said that in one month of production traffic, Claude and GPT requests through Perplexity had higher success rates than any individual provider measured, with less daily variation during provider incidents. It said that for teams running millions of requests a day, even the difference between 99.0% and 99.99% success can affect thousands of responses. However, it did not give the success rates it compared, the measurement period or the names of the providers compared. The measurement was Perplexity's own, and ORIENS could not confirm any third-party verification.
Automatic model selection is in preview
"Routes," which picks models automatically, and "Auto effort," which adjusts reasoning depth, are in preview. With Routes, teams describe in plain language which kinds of requests go to which model and set a default model for when none fits. Teams can start from built-in "Lower cost," "Balanced" or "Higher quality" profiles or build their own.
According to the documentation, specifying perplexity/auto as the model has a decision model read the conversation and each route's description to pick a model. One to five routes can be defined. Under the default policy without custom routes, GPT-6.1 Sol answers, while greetings, short translations and similar requests go to GPT-6 Luna. The default model answers when the decision model fails or times out, or when the conversation contains images, audio, video or files. Billing is at the rate of the model that actually answered.
With Auto effort, the decision model picks low, medium or high reasoning effort. The blog cited an agent spending more effort to recover from a tool error without switching models. Perplexity did not say when the preview will end or when general availability will begin.
Prices of major models
Prices are in U.S. dollars per 1 million tokens, set separately for input and output. GPT-6.1 Sol, the main model under the default automatic selection, costs $2.00 for input and $10.00 for output (for inputs up to 272,000 tokens), and GPT-6 Luna, which handles light tasks, costs $0.10 for input and $0.50 for output (same condition). Claude Opus 5.5 costs $4.00 for input and $20.00 for output, the open-weight GLM-5.3 costs $1.40 and $4.40, and DeepSeek V4 Pro costs $1.32 and $3.96. The Gemini 3.8 Flash rates of $0.75 for input and $3.75 for output apply through Dec. 31.






