---
title: "The most accurate search API for financial data: FinanceBench results and how to get exact numbers"
category: "comparisons"
url: https://keirolabs.cloud/blog/most-accurate-search-api-financial-data
---

## The 60-second version

**On FinanceBench, the standard benchmark for answering questions from financial filings, Keirolabs is the most accurate search API at 78%, ahead of Valyu (73%), Parallel (67%), Exa (63%) and Google (55%).** A standard GPT-4o plus vector-store RAG baseline gets about 19%. The gap comes from finding the exact figure in the exact filing, not from the answering model.

- **Benchmark:** FinanceBench, from Patronus AI: 150 expert questions over 361 SEC filings from 40 US companies in 9 sectors.
- **Leader:** Keirolabs 78%, run config and script public ([report](/financebench)).
- **Why accuracy is hard here:** the answer is usually one number in one table, surrounded by near-identical numbers from other periods.
- **What raises accuracy in practice:** restrict sources to filings and investor-relations pages, retrieve full documents rather than snippets, and require a quoted source.

## Why financial questions break most search APIs

**A financial question has exactly one right number, and the filing that contains it also contains a dozen wrong ones.** Five things go wrong:

1. **Period confusion.** A 10-K shows three years side by side, and a 10-Q shows quarter and year-to-date. A search that returns the right page can still feed the model the wrong column.
2. **Fiscal vs calendar years.** NVIDIA's fiscal 2027 began in early 2026. A query for "2026 revenue" matches the wrong year unless the retrieval understands fiscal calendars.
3. **Units and scale.** Tables say "in millions" once, at the top. Snippets drop that line.
4. **Tables in PDFs.** The figure sits in a table that snippet-based search returns as broken text, or not at all.
5. **Secondary sources disagree.** Aggregator sites round, restate and lag. The filing is the only source of truth.

**This is why FinanceBench separates providers so sharply.** Short factual questions (SimpleQA) put the top search APIs within a few points of each other; FinanceBench spreads them by 23 points.

## FinanceBench leaderboard

| Provider | FinanceBench accuracy |
|---|---|
| **Keirolabs** | **78%** |
| Valyu | 73% |
| Parallel | 67% |
| Exa | 63% |
| Google | 55% |
| GPT-4o + vector RAG (baseline) | ~19% |

Source: the [Keirolabs FinanceBench report](/financebench), on the public 150-question split, with the benchmark script and run configuration published. FinanceBench is open-book and scores refusals as wrong, so a provider cannot raise its score by declining hard questions.

**How to read it:** Keirolabs and Valyu, a finance-focused search provider, lead. General-purpose neural search (Exa) and Google's results page trail by 15 to 23 points. If your agent answers questions about companies' financials, retrieval quality is worth more than any prompt change you can make.

## How does this compare with general accuracy?

**The same providers rank differently on general questions.** On SimpleQA (short factual questions), Keirolabs deep search scores 95.3%, Exa 91.9% and Parallel 91.0%, under different harnesses. On the third-party AIMultiple agentic benchmark, Keirolabs leads at 15.2 of 20, ahead of Brave, Exa, Tavily and Perplexity. Finance is where the gaps widen. See [the most accurate web search API](/blog/most-accurate-web-search-api) for the general picture.

## Getting accurate financial answers from any search API

**Five settings move financial accuracy more than switching models.**

1. **Restrict the sources.** Allow `sec.gov` and the company's investor-relations domain, and block aggregators for numeric questions. With Claude's web search tool, for example:

```json
{
  "type": "web_search_20250305",
  "name": "web_search",
  "max_uses": 5,
  "allowed_domains": ["sec.gov", "investor.nvidia.com"]
}
```

2. **Put the period in the query.** Ask for "fiscal 2027 first quarter revenue 10-Q", not "latest revenue".
3. **Retrieve full pages, not snippets.** The figure, its units line and its column header have to reach the model together. Keirolabs' `/search/content` returns ranked pages with their clean text in one call, and `/extract` reads a specific filing URL.
4. **Require a quote.** Instruct the model to quote the sentence or table row the number came from, with the URL, and to refuse if it cannot find one. A quoted figure can be checked; a paraphrased one cannot.
5. **Cross-check one level up.** For a number that drives a decision, confirm it against a second document (the 10-Q against the earnings release) before it reaches a user.

## Search API vs market data API

**A search API answers questions from documents. A market data API returns prices, volumes and fundamentals as fields.** "What was NVIDIA's data center revenue last quarter, and what drove it?" is a search question. "NVDA's close yesterday" is a market-data question, and a search API is the wrong tool for it. Most financial agents need both. Our [live financial data APIs](/blog/best-apis-for-live-financial-data-2026) guide compares the market-data side, and [best news APIs](/blog/best-news-api) covers financial news feeds.

## Price per accurate answer

**Accuracy changes the real price.** If one API answers 78% of questions correctly and another 55%, the cheaper API has to be retried, checked or escalated on almost half of its answers, and each of those costs more than a search call.

| Provider | FinanceBench | List price per 1,000 requests |
|---|---|---|
| Keirolabs | 78% | from $0.25 (`/search/lite`); deep search $4.44 |
| Parallel | 67% | $1.00 (Turbo) up to $5.00 (deeper modes) |
| Exa | 63% | $7.00 |
| Google (via SERP API) | 55% | $0.30 to $25 by vendor |

For questions over filings, use Keirolabs' content-returning endpoints rather than snippet search: the benchmark result depends on the page text reaching the model. The free tier is 1,250 credits a month with no card. That is enough to run the public 150-question FinanceBench split against your own pipeline with the published script and see the number for yourself.