Parallel Web Systems Beats OpenAI and Exa on Finance Agent Search
Parallel rolls out finance-tuned web search for AI agents, posting top accuracy on three benchmarks while undercutting Exa, Perplexity and OpenAI on cost.
- Parallel launched Parallel for Finance, a universal index upgrade applied across all its APIs with no new subscription.
- On FinSearchComp T3, Parallel Advanced hits 57.4% accuracy at up to 2x lower cost than Exa and Perplexity.
- Parallel Fast scores 53.7% on T3 at $15.2 per 1,000 questions, roughly 10x cheaper than rivals.
- On Samaya's FrontierFinance, Parallel ties for top at 38% lower cost than Exa.
- On its private PWS-Finance benchmark, Parallel beats OpenAI's native search by 30% in cost.
- Rogo, Kepler, Bretton and Taktile already run Parallel in production for finance workflows.
Parallel Web Systems has launched Parallel for Finance, an upgrade to its web index and ranking systems for agents that search filings, regulator registers, exchange portals, company disclosures, and financial news. Existing API users receive the changes automatically under the company’s current products and pricing.
Finance agents can produce incorrect answers by retrieving the wrong fiscal period, currency, filing version, or corporate entity. Resolving those ambiguities often requires primary documents beyond a team’s licensed datasets. Parallel argues that ranking those sources more effectively can improve accuracy while reducing the number of searches, fetches, and model tokens required for each answer.
One harness, three tests
Parallel evaluated its search against Exa, Perplexity, OpenAI native search, and Tavily using two external benchmarks and one internal test set. Every provider ran through the same agent harness, a setup designed to isolate the effect of the retrieval service. Parallel conducted and reported all three evaluations.
FinSearchComp tests financial research across global and Greater China markets. Its T3 tier contains complex historical questions that require several retrieval and reasoning steps. Parallel Advanced recorded the highest accuracy in the company’s run, while Parallel Fast outscored every listed competitor at a substantially lower cost.
The cost metric is CPM, measured in US dollars per 1,000 completed questions. It includes answer-model tokens and all search and fetch calls at list prices.
| Provider | Accuracy | Cost per 1,000 questions |
|---|---|---|
| Parallel Advanced | 57.4% | $161.10 |
| Parallel Fast | 53.7% | $15.20 |
| Exa Auto | 51.4% | $226.20 |
| Perplexity Finance + Web | 50.0% | $347.00 |
| OpenAI native search | 48.3% | $204.30 |
| Tavily | 43.0% | $481.80 |
Parallel Fast exceeded the nearest competitor, Exa Auto, by 2.3 percentage points while costing about 93% less. Advanced added 3.7 percentage points over Fast but cost more than ten times as much, giving developers two distinct accuracy and budget profiles.
In Parallel’s run of Samaya’s FrontierFinance benchmark, which grades long-form research answers with three LLM judges and task-specific rubrics, Parallel tied for the highest score at a reported cost 38% below Exa. The company also reports that it beat OpenAI, Perplexity, and Tavily on both accuracy and cost.
The in-house test goes wider
Parallel’s most tailored evidence comes from PWS-Finance, its held-out internal benchmark. Vendor-owned tests give the author control over question selection, answer labels, and evaluation settings, so Parallel documented several measures intended to reduce bias. The company says it sourced questions from real research demand, used each competitor’s highest-capability agent mode to help construct gold-standard answers, and followed each provider’s published guidance within a common harness.
The 174 questions ask for specific facts about a single entity as of a defined date. They cover eight common workflows:
- Company profiles
- Funding and investors
- Deals and transactions
- Ownership and leadership
- Public financials
- Regulation and compliance
- Markets and macroeconomics
- Recent developments
Geographic coverage includes 65 questions about the United States, 34 about Europe, and 24 about Greater China. The remaining 51 span Asia-Pacific, the Middle East, Africa, and Latin America.
Most questions include a condition that can expose shallow retrieval. Relevant facts may be divided among several sources, published months after the period they describe, or changed between filing versions. Parallel cross-checked answer keys with multiple research agents, critic models, and a verifier. In a separate validation exercise, that process detected 93% to 100% of deliberately inserted errors.
Each response receives credit only when every required fact is correct. Under that strict scoring method, Parallel Advanced ranked first while costing 30% less than OpenAI native search. Its score exceeded Exa and Perplexity by 8 to 10 percentage points. Parallel Fast finished within one percentage point of OpenAI at a reported cost 39% lower.
What changes in an agent stack
For developers building research, diligence, or monitoring agents, the results identify retrieval as a measurable source of both answer quality and operating cost. A stronger index can reduce failed searches and unnecessary model calls, while better ranking can place the relevant filing or disclosure earlier in the context window.
Teams comparing providers can reproduce the evaluation pattern with their own workloads:
- Build a held-out set of representative questions whose answers are absent from public benchmark pages.
- Grade dates, currencies, entities, filing versions, and citations as separate fields.
- Track total model tokens, search calls, fetch calls, latency, and list-price cost.
- Inspect date filters, call limits, and fallback behavior for difficult queries.
- Compare fast and high-accuracy modes against the same acceptance threshold.
Parallel identifies Rogo, Kepler, Bretton, and Taktile as finance customers. Rogo builds AI research tools for investment banks, and CEO Gabriel Stengel says its agents handle hundreds of thousands of requests on peak days across filings, earnings calls, and market sources. Parallel’s post does not provide production latency, failure-rate, or per-request cost data for those deployments.
The company also released an open-source tear-sheet cookbook. The Python recipe accepts a CSV of stock tickers and generates two-page equity reports containing four years of financials, market data, recent news, bull and bear cases, citations, and a research-basis page.
Where the rankings can move
-
Published answers affect FinSearchComp. Answers to the benchmark are available online. Parallel removed those pages from results returned by the other search APIs before the answer model saw them. For OpenAI native search, Parallel marked answers sourced from published copies as incorrect. That policy reduced OpenAI’s T3 score from 64.0% to 48.3%, changing it from the highest unadjusted score to fourth place in the reported table.
-
Tavily is absent from PWS-Finance. Parallel excluded Tavily after the model applied a one-year date filter to 98% of searches and reached the 20-call limit on 77% of questions. The outcome may reflect interactions among the harness, model-generated tool calls, and provider limits, and it leaves no Tavily comparison for the internal benchmark.
-
Parallel remains the sole reporter. The external datasets provide common questions, but Parallel selected the harness, prompts, provider settings, pricing assumptions, and scoring implementation. Results may change with another answer model, a different query mix, negotiated pricing, caching, or updated search indexes.
The published evidence is strongest for dated, auditable financial questions that require primary-source retrieval. Parallel’s results make its finance-tuned index a candidate for workload-specific testing, especially where search and fetch costs dominate. Applying the same approach to legal, life-sciences, or insurance research remains an extrapolation until comparable benchmarks cover those domains.
Comments
