Exa vs Parallel
What both are asked for. Given a company domain, return that company's latest funding stage: the most recent round it raised, named correctly. Both receive the same domains and the same output schema, and every answer is judged against a human labelled dataset.
Two boards. Freshness measures rounds announced in the last 30 days: how fast a provider indexes news. Enrichment measures rounds older than that: how completely it has backfilled history. These reward opposite things, so they are ranked separately. Each freshness cohort ages into enrichment once its window closes.
What was called. Exa through POST /agent/runs (effort=medium · JSON schema), POST /search (type=deep-reasoning · JSON schema), POST /search (type=instant · JSON schema). Parallel through POST /v1/tasks/runs (processor=core · structured JSON output), POST /v1/responses (model=parallel · reasoning.effort=medium · text.format JSON schema). Where a vendor exposes more than one, each endpoint is measured on its own and takes its own row on the boards below.
How Exa and Parallel scored on recent rounds
Rounds announced inside the window, with the cohort rebuilt each cycle. Each endpoint is measured on its own and takes its own row, so the configuration column is where a vendor's arms separate. This is the rolling freshness board, pooled across every dated snapshot, so each row carries the number of companies that provider was measured on. The company funding benchmark has other vendors that were benchmarked on the same input data. The methodology covers cohort design, Ground Truth and the judging rules.
| Provider | Endpoint & configuration | Measured on | Latest stage correct | Correct when returned | Latest stage returned | Funding fields returned | Fields returned | Median latency | Est. cost | Official docs | Company identified |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Long Running Agent APIs | |||||||||||
| Exa | POST /agent/runseffort=medium · JSON schema | 100company measurements · 2 snapshots | 97.0% | 97.9% | 96.0% | 73.6% | 3.7 / 5 | 27,293 ms | $10.00$0.10 / request (medium effort) | Exa official docs | 100.0%100/100 domains |
| Parallel | POST /v1/tasks/runsprocessor=core · structured JSON output | 181company measurements · 3 snapshots | 95.0% | 96.0% | 95.6% | 86.7% | 4.3 / 5 | 62,142 ms | $4.53$25 / 1,000 Task runs | Parallel official docs | 100.0%181/181 domains |
| Web search APIs | |||||||||||
| Exa | POST /searchtype=instant · JSON schema | 100company measurements · 2 snapshots | 98.0% | 99.0% | 95.0% | 90.2% | 4.5 / 5 | 1,618 ms | $0.70$0.007 / request | Exa official docs | 100.0%100/100 domains |
| Exa | POST /searchtype=deep-reasoning · JSON schema | 181company measurements · 3 snapshots | 95.6% | 96.0% | 96.7% | 84.8% | 4.2 / 5 | 9,916 ms | $2.71$0.015 / request | Exa official docs | 100.0%181/181 domains |
| Parallel | POST /v1/responsesmodel=parallel · reasoning.effort=medium · text.format JSON schema | 100company measurements · 2 snapshots | 90.0% | 90.6% | 96.0% | 82.6% | 4.1 / 5 | 17,020 ms | $5.00$50 / 1,000 Responses requests (medium reasoning) | Parallel official docs | 100.0%100/100 domains |
How Exa and Parallel scored on settled rounds
The same vendors on rounds that were already more than 30 days old when each one ran. Read the company identified column before the rates: arms joined the benchmark at different points, so they do not all sit on the same number of companies. This is the funding enrichment board, scored on rounds that were already more than 30 days old when each provider ran. The company funding benchmark also carries the rolling freshness board. The methodology covers cohort design, Ground Truth and the judging rules.
| Provider | Endpoint & configuration | Latest stage correct | Correct when returned | Latest stage returned | Funding fields returned | Fields returned | Median latency | Est. cost | Official docs | Company identified |
|---|---|---|---|---|---|---|---|---|---|---|
| Long Running Agent APIs | ||||||||||
| Exa | POST /agent/runseffort=medium · JSON schema | 88.7% | 93.0% | 90.7% | 67.3% | 3.4 / 5 | 26,549 ms | $30.00$0.10 / request (medium effort) | Exa official docs | 100.0%300/300 domains |
| Parallel | POST /v1/tasks/runsprocessor=core · structured JSON output | 90.0% | 92.3% | 95.4% | 84.9% | 4.3 / 5 | 53,684 ms | $5.47$25 / 1,000 Task runs | Parallel official docs | 100.0%219/219 domains |
| Web search APIs | ||||||||||
| Exa | POST /searchtype=deep-reasoning · JSON schema | 88.6% | 88.6% | 100.0% | 93.6% | 4.7 / 5 | 12,848 ms | $3.29$0.015 / request | Exa official docs | 100.0%219/219 domains |
| Exa | POST /searchtype=instant · JSON schema | 87.0% | 89.0% | 97.0% | 89.0% | 4.5 / 5 | 1,507 ms | $2.10$0.007 / request | Exa official docs | 100.0%300/300 domains |
| Parallel | POST /v1/responsesmodel=parallel · reasoning.effort=medium · text.format JSON schema | 89.0% | 90.1% | 94.3% | 81.6% | 4.1 / 5 | 15,930 ms | $15.00$50 / 1,000 Responses requests (medium reasoning) | Parallel official docs | 100.0%300/300 domains |
Contact data, intent, and integrations are not measured here. This is the company-funding slice only. Full leaderboard, methodology, and per-company evidence →
Where each one fits
- Enriching a known list. Speed and accuracy both matter here. Web Search APIs and long running APIs both have comparable accuracy, but the Web Search APIs leads on speed.
- Inside a request path. An API that takes minutes cannot sit in a blocking call. If the enrichment has to finish before something else runs, the web search APIs lead here.
- Real time funding data. Accuracy is the only metric that matters for fresh funding data. Both the vendors have comparable scores, so pick on price and latency based on the use case.
Exa vs Parallel: common questions
Does an agent beat a search index on funding data?
Not automatically, and the two boards disagree. On rounds announced in the last month a good index tends to have the answer already, because the announcement was news and got crawled. On older rounds the gap widens, since an agent can chase a source that no index surfaced.
Are these two sent the same request?
The same instruction and the same output schema, yes. That is deliberate: it is what makes the difference in score attributable to the endpoint rather than to a differently worded question. Contract tests in the open repository assert the two payloads match apart from the endpoint.
Which one should a batch pipeline use?
Batch removes the agent's main disadvantage, because a slow answer costs nothing when nothing is waiting on it. At that point the decision is cost per company against how many more companies you need answered correctly.
Can a result from either be traced back to a source?
Yes for both, and that is part of why they are compared. Each was asked for the stage together with the material it relied on, so a returned answer can be opened and checked against the announcement it came from. A database row that arrives with no provenance cannot be audited the same way.