Which API fits the job?
Exa's lower medians in our August fixture set do not establish that every Exa mode beats every Parallel mode. We tested Parallel advanced, not its newer low-latency choices. First decide whether the workflow needs a compact source lookup or richer objective-led excerpts.
When to consider Exa
Consider Exa when search and known-URL contents are separate steps, or when the lower local medians in our fixture set match your latency priorities. Validate date handling before assuming every result is timestamped.
When to consider Parallel
Consider Parallel when you want to express the research objective alongside search queries and keep excerpts inside an existing agent loop. Its current turbo/fast price is a different option from the advanced mode we measured.
Our measured retrieval results
August 23, 2026. Three developer-source tasks, three requests each per provider. Every expected source ranked first in both runs. These are historical configuration-specific results, not a new mode comparison.
| Task | Exa median | Parallel median |
|---|---|---|
| Find the current Node.js release lines and identify which releases are LTS. | 145 ms | 2,044 ms |
| Find the current GitHub-hosted Actions runner images and their included software. | 130 ms | 1,732 ms |
| Find the PostgreSQL versions currently supported by the project and its support policy. | 171 ms | 2,040 ms |
Date-field coverage was 33% for Exa and 49% for Parallel; snippet coverage was 100% and 100%. Neither field coverage nor local speed alone measures final-answer quality.
Exa exact queries and artifacts · Parallel exact queries and artifacts
Our fixtures do not establish global latency, production uptime, consumer-answer fidelity or Google rankings. Account setup and starting credits also describe the original run, not a guaranteed current signup offer.
Request shape, pricing and agent integration
| Decision | Exa | Parallel |
|---|---|---|
| Request contract | Query with an explicit retrieval mode and content choices | Objective and/or search queries with mode and excerpt settings |
| Current lower-cost mode | Lite / Instant: $4 per 1,000 requests, up to ten results | Turbo / Fast: $1 per 1,000 requests, default ten results |
| Other published retrieval modes | Auto / Fast: $7 per 1,000 requests, up to ten results | Basic / Advanced: $5 per 1,000 requests, default ten results |
| Content beyond a result list | Contents for supplied URLs; content options and extra results can add usage | Extract for supplied URLs; additional page results/excerpts add usage |
| Agent connection | Official MCP plus SDKs and REST | Official MCP, CLI, SDKs and REST |
Prices are published USD list rates or explicitly labeled credit calculations, not observed costs from a new test. Extra results, content options, model tokens, account limits and minimum commitments can change the total. Check the linked source for the exact configuration.
A useful evaluation before switching
- Run one fixed documentation lookup in each exact mode you might ship; do not compare Exa Instant with Parallel Advanced and label it a provider-wide race.
- Measure how many returned excerpts the agent can use without another fetch. A slightly slower first response may still finish the task sooner.
- Normalize optional publication dates without inventing them. Compare downstream completion before assigning a weight to date coverage.
Keep the same acceptance criteria across providers and save failures as failures. Avoid changing the target source, query or score threshold after seeing which vendor wins. Measure both useful evidence and the total workflow needed to obtain it.
Sources and review dates
Documentation review: October 27, 2026. Our own measured retrieval runs: August 23, 2026. External reports retain their publisher and measurement window.
Corrections: email faizank@docsalot.dev with an official source.