Agent Bench

Exa vs Perplexity Search API

Both passed all nine source lookups. Exa had lower local latency; Perplexity returned more complete date metadata. Compare the results alongside current search modes and what 10,000 requests cost.

Tested Aug 23, 2026 18 total requests Pricing reviewed Nov 20, 2026

Choose by retrieval mode, source dates and workflow.

Consider Exa when

You want query-led discovery with selectable content options, or the lower latency observed in these three cases. Lite uses cached content; use a suitable mode if you need a live fetch.

Consider Perplexity when

Your workflow needs the date fields that were complete in this run, or already uses Perplexity billing. Fast Search has a lower published rate than Exa Lite, but these newer modes were not tested head-to-head.

Exa and Perplexity results

These are August observations, not a retest of current modes. Latency is from one machine and network. The small developer-documentation fixture set does not establish general relevance or a production SLA.

MeasureExaPerplexity Search
Attempts passed9/99/9
Primary sourceRank 1 in all 9 attemptsRank 1 in all 9 attempts
Local median range130-171 ms427-1,045 ms
Snippet coverage100%100%
Date metadata33%100%
Average results1010
Payment cardNo credit card requiredCard required for a new API group
Starting credit observed in August$10 during our signupNot measured on a clean account

Starting credits and card requirements describe our setup at the time. Perplexity recovered an existing organization, so its clean-account signup and starting balance were not measured.

What did we ask each API?

Each task ran three times per provider. The expected primary source was rank one in all attempts on both APIs. Values below are local median request times, not model-answer latency.

TaskExpected sourceExa medianPerplexity median
Find the current Node.js release lines and identify which releases are LTS.nodejs.org145 ms427 ms
Find the current GitHub-hosted Actions runner images and their included software.github.com or docs.github.com130 ms436 ms
Find the PostgreSQL versions currently supported by the project and its support policy.postgresql.org171 ms1,045 ms

A source ranked first is not the same as a correct generated answer. We did not evaluate consumer-site answers, geographic rank fidelity, broad topic coverage or a statistically reliable speed difference. The linked reports preserve the individual runs and queries.

Pricing depends on the search mode

Published USD prices checked November 20, 2026. These are not measured run costs or equivalent-quality tiers. The newer modes below have not been benchmarked here.

ModePer 1,000 requestsBudget caveat
Exa Lite$4Up to 10 results; cached content only. Instant is a legacy mode.
Exa Auto$7Up to 10 results; extra results and summaries can add cost. Fast is a legacy mode.
Perplexity standard Search$5Successful requests, including empty results. No extra extraction-token charge.
Perplexity Fast Search$1A separate mode; our August run does not establish its retrieval quality.

What would 10,000 searches cost?

Before free credits and add-ons: Exa Lite $40, Exa Auto $70, Perplexity Fast $10, or Perplexity standard Search $50. Exa assumes at most ten results per request. Perplexity bills one successful request even if it contains multiple queries, up to five. Compare the same query workload before treating those units as interchangeable.

Recheck Exa pricing and Perplexity pricing against the exact request you plan to send. A request budget should include agent loops, retries, extraction and any model used to write an answer.

Compare the same job, not just the brand

Start with the output your application needs. A ranked source list, an extracted page and a generated answer need different acceptance tests.

Does Perplexity Search write an answer?

No. This comparison uses its ranked-results endpoint. For a written answer, consult the current Agent API documentation; older Sonar tutorials may describe a different API. See our Perplexity API integration guide for the endpoint distinction.

Can a coding agent use either through MCP?

Yes, both now document official integrations: Exa MCP and Perplexity MCP. These retrieval measurements used API calls, not a head-to-head MCP or coding-agent tool-use test.

Which is better for scraping?

For known-URL reading, compare Exa Contents with Perplexity's separately documented Context API; neither was tested here. Capturing a consumer-site answer is a different job. Our Perplexity scraper vs Search API guide separates those requirements.

What should I test before switching?

Use a handful of queries from your application and name an acceptable source for each. Save the parameters, source order, missing fields, failures and cost. Repeat them across your actual workload before treating our nine attempts as a provider-wide verdict.

Inspect the runs before choosing.

Each report includes the exact queries, ranked results, per-case latency, pass criteria, onboarding record, caveats, and downloadable JSON.

All search APIs Exa vs ParallelParallel vs PerplexityExa vs TavilyTavily vs BraveExa vs FirecrawlExa benchmark Perplexity benchmark Test methodology