Perplexity Decider API pricing

v1.1 and v1 have the same published rate

The official pricing table lists both versions separately. The current quickstart uses v1.1 and requires an explicit model on each request. Equal prices do not mean identical behavior. Keep the version in your evaluation records so an upgrade can be checked against the same labeled examples.

These rates apply to the Decisions API. Sonar search, the Router API and other Perplexity services have their own pricing. A Perplexity consumer subscription should not be treated as a fixed allowance of these API calls.

What the endpoint buys

Decider returns probabilities for questions about a supplied state. It can answer a yes/no question with noul, choose among your options with choice, or score content against ordered levels with score. Your application uses those numbers to decide what to do next.

For example, a support system could ask whether a message needs human review, then choose a routing category. A content workflow could check an image against a stated criterion. The service evaluates the supplied content; it does not generate a chat reply, write an explanation or create an image.

Probability is not a guarantee of correctness. Define an escalation threshold using labeled examples and the cost of a wrong decision. A 0.9 output should not automatically become a business-wide threshold without calibration on your own workload.

Calculate cost from billable input

The billable count includes the state, every question and any images. Read usage.input_tokens from the response, then calculate cost_usd = usage.input_tokens × 0.02 ÷ 1,000,000. Avoid estimating from the state text alone when you attach long criteria or many questions.

The 600-token example is also given in Perplexity's pricing documentation. The volume examples above assume that each call has the stated complete billable input, with no retries. They are arithmetic scenarios, not invoices from a production application.

Suppose 100,000 routing requests contain 1,200 billable tokens each after the questions are included. The decision stage costs $2.40. If 5% escalate to another service at an assumed $0.01 per call, add $50, for a total of $52.40. The downstream rate and escalation percentage are assumptions; replace both with your own measurements.

Multiple questions and images are not free input

No request fee means you do not add a flat charge for each HTTP request. It does not mean an unlimited questionnaire has the same cost. Every added question increases the input that the service must read. If you combine questions about a shared state, compare the reported usage with the alternative of sending that state repeatedly.

Images also consume input tokens. Use the reported token count rather than assigning a universal price to a photo. Resizing an image or changing the amount of accompanying text can change usage. Output probabilities are free under the published rate; the image and questions remain billable.

Request size and throughput are different limits

The quickstart allows 1–128 named questions and requires total input to be under 262,144 tokens. It documents a 32 MiB request-body limit and image-size limits. Remote image URLs are not accepted; use a supported image data URL and follow the documented size guidance.

A request can fit the size limits yet still exceed your calling rate. Check Perplexity's current rate-limit guidance before sizing concurrent workers. Adding a queue or delaying work changes throughput; it does not change the $0.02 unit price. Retries that become additional billable calls belong in the usage total.

Compare with OpenAI Decisions without mixing endpoints

Our OpenAI Decisions API pricing guide explains GPT-6 Luna's separate decision endpoint, including long-context and regional exceptions. Compare endpoint-specific prices and the input each system actually consumes. The general LLM calculator models text generation and is not a Decider billing calculator.

A useful evaluation records the model version, exact question criteria, total billed input, accepted decisions, wrong decisions and escalations. Keep quality and latency alongside spend. Comparing one provider's input rate with another provider's input-plus-output chat bill can answer a budget question, but it does not prove the two services solve the same task equally well.

To start an estimate, count a representative request using the documented usage field, multiply by planned volume, then add retries and downstream actions. Recheck the first-party price table before committing to volume. AI API Prices editors verified this page on October 9, 2026; corrections can be sent to [email protected].