We benchmarked three search APIs for coding agents
Exa, Perplexity Search, and Parallel Search all passed our retrieval test. The useful differences appeared in latency, metadata, onboarding, and integration surface.
An agent that searches the web needs more than a list of vaguely related links. It needs to find the primary source, preserve enough context to judge the result, and return predictable fields that the next tool call can use.
We tested Exa, Perplexity Search, and Parallel Search against the same three research jobs. Each provider received nine requests: three fixed queries, repeated three times. Across all 27 attempts, every provider returned the expected primary source at rank one.
That tie is the interesting part. A simple pass count would make the providers look interchangeable. They were not.
What the test measured
The benchmark asked each API to find current first-party information about:
- Node.js release lines and LTS status
- GitHub-hosted Actions runner images
- PostgreSQL supported versions and support policy
For an attempt to pass, the API had to return at least five valid results, provide snippets for at least 80% of them, and place an expected primary source in the top five. We requested ten results with provider-default geography and a 30-second timeout.
You can inspect the versioned protocol, the exact queries, and all normalized result artifacts. Credentials, response headers, and account identifiers are excluded from the public files.
The result at a glance
| Provider | Attempts passed | Primary source | Median latency across tasks | Date coverage |
|---|---|---|---|---|
| Exa | 9/9 | Rank 1 every time | 130–171 ms | 33% |
| Perplexity Search | 9/9 | Rank 1 every time | 427–1,045 ms | 100% |
| Parallel Search | 9/9 | Rank 1 every time | 1,732–2,044 ms | 49% |
These are observations from one controlled run on one machine. They are not universal reliability or global latency claims.
Exa was fastest in this run
Exa recorded the lowest median latency for all three jobs. Its medians ranged from 130 to 171 milliseconds, while still returning ten results per request and placing the expected primary source first every time.
The tradeoff appeared in date metadata. Only 33% of Exa's normalized results included a publication date. That may be irrelevant for stable documentation, but it matters when an agent must distinguish a current release note from an older page.
Read the Exa benchmark report for the query-level results and downloadable artifacts.
Perplexity returned the most complete dates
Perplexity Search was the only provider in this run with publication dates on every normalized result. It also passed all nine attempts with the primary source ranked first.
Its median latency varied more by task, from 427 milliseconds for the Node.js query to 1,045 milliseconds for PostgreSQL support. The credential came from an existing API organization, so our report does not claim anything about new-account starting credit or signup friction.
Read the Perplexity Search report for the full evidence.
Parallel exposed the broadest agent-facing path
Parallel Search also passed all nine attempts with stable source ordering. The observed medians were slower than the other two providers in this run, ranging from 1,732 to 2,044 milliseconds.
Its product surface is unusually explicit about agent use: the API is available alongside SDK, CLI, MCP, and coding-agent integration paths. We observed a starting balance during account setup and recorded the amount, but that is evidence from one account rather than a promise for future signups.
Read the Parallel Search report for the onboarding notes and result bundle.
What this benchmark does not answer
This release isolates retrieval. It tells us whether a provider can return useful, source-ranked search results for fixed technical questions. It does not prove that a coding agent can discover the product, create an account, choose the right endpoint, manage authentication, or integrate the API into a repository without help.
Those are separate agent-experience questions, and they deserve a separate protocol. Keeping them apart prevents a fast endpoint from being mistaken for an easy integration, or polished documentation from being mistaken for good retrieval.
For the current evidence, start with the Agent Bench report library. Each provider report includes the exact inputs, representative outputs, measured fields, setup observations, and a sanitized artifact download.