Perplexity Fast Search: $1 per thousand requests, configuration and billing

Search results are not a complete answer

Perplexity added Fast Search October 5, 2026 as a Search API search type retrieving web results for your software. One call does not complete a reasoning, citation and action-heavy agent task.

The documented response format stays the same, so existing parsers should focus on the requested type. For generated answers with sources, check the appropriate generation endpoint and charges.

Fast Search's fast is a search type; Agent API fast is another preset layer. Changing only a preset cannot establish that direct Search API requests use the $1 tier.

Understand the successful-request billing unit

Current prices are $5 Standard and $1 Fast per thousand requests. The unit is one successful POST /search, not number of pages returned or keywords.

Up to five queries in one successful request count once. Successful empty responses charge; invalid, rate-limited and upstream-failed requests do not. Check success before result count.

Pure arithmetic example: 100,000 successful requests cost $100 Fast versus $500 Standard. No calls were executed and generation, other tools and your services are excluded.

Search API itself adds no token charges. Subsequent model processing has its own cost; the search price is not a full complex-research-task price.

Existing calls do not switch automatically

First confirm /search and a valid key, kept server-side or in secure local configuration. Field examples are not credentials for public pages.

Second explicitly set search_type to fast. Omission defaults to Standard web search; release alone does not move old requests into the new tier.

Third follow current official required fields such as query and validate a parseable successful response. max_results accepts 1–20; increasing it does not select fast.

Fourth inspect Fast Search API Requests Count in console Search API usage against actual requests. Web membership or an entire generation task's bill is not this metric.

SDK parameter errors happen before networking

For Python 0.43.4/0.43.5, docs identify validation trouble with direct search_type="fast" and recommend passing it through extra_body.

For TypeScript 0.38.5 they show as any on the object for parameter typing, not disabling project-wide checking. Consult the linked official full example.

These conditions concern those versions, not all future Python SDKs forever. Pre-request argument errors are not proof of service failure, changed balance or exit problems.

Retain a search choice for difficult questions

Fast targets routine agent loops, frequent retrieval and responsiveness. Docs recommend Standard web for difficult, obscure or ambiguous questions. This is positioning, not our speed/accuracy test.

Hypothetical unexecuted example: retrieve an official product page before extracting a release date. Compare whether fast returns the necessary original; if ambiguity makes results hard to judge, choose Standard explicitly.

Record whether sources answer the question, their reliability and each stage's actual cost. Search success establishes a response, not verified facts; open originals for important conclusions rather than relying on ranking.