Nov 6, 2026
- [Parallel Decisions is now on OpenRouter](https://parallel.ai/blog/decisions-openrouter)
Tags:Product
Author: By ParallelNovember 6, 2026
# Parallel for Legal: Setting the new standard for Legal AI Search
Tags:Product
Today we’re launching **Parallel for Legal**, an upgrade to our search index and ranking built for legal agents. With this upgrade, Parallel finds hard-to-reach primary sources across jurisdictions and returns trustworthy evidence agents can cite. Parallel for Legal is available across all Parallel APIs, with no new plan or subscription required.
Legal AI agents need current and authoritative sources to answer accurately. However the critical evidence that is needed is scattered across multiple different sources including court dockets, state codes, agency guidance, and government publications, and more. Finding and extracting the correct responses often requires teams to build and maintain their own crawling pipelines.
Key takeaways:
- - On PWS-Legal, Parallel’s custom benchmark built around legal questions, Parallel leads in accuracy and holds the entire Pareto frontier, **beating every other provider at up to 2.3× lower cost**.
- - On LegalBench Citation[LegalBench Citation], Parallel leads every other search API on accuracy, at the same or lower cost. On Legal Hallucinations[Legal Hallucinations], Parallel matches OpenAI and Exa’s accuracy at **2.6× lower cost**.
- - Leading legal AI companies, including Harvey AI[Harvey AI], GC AI[GC AI], and Finch[Finch] already use Parallel in production.
## Defining the cost-accuracy frontier for legal
We tested three legal benchmarks with the same agent, questions, and grader, changing only the search provider on each of the runs. The cost metric includes model tokens, search, and fetch at list price. You can find more details on the methodology below each chart.
PWS-LegalLegal HallucinationsLegalBench Citation
Correct answers per $
Parallel Advanced4.3/$ | 83.1% / 194.2CPM
Parallel Fast3.7/$ | 85.4% / 230.1CPM
Perplexity3/$ | 73.4% / 248CPM
Firecrawl2.5/$ | 77.3% / 314.4CPM
Exa Auto2.4/$ | 75.6% / 318.9CPM
OpenAI1.8/$ | 80.5% / 440.8CPM
Tavily1.8/$ | 75.4% / 425.7CPM
COST (CPM)
ACCURACY (%)
Loading chart...
CPM: USD per 1,000 questions, end to end (answer-model tokens plus search and fetch calls at list price).
Parallel
Others
Benchmark comparison across Cost (CPM) and Accuracy (%). CPM: USD per 1,000 questions, end to end (answer-model tokens plus search and fetch calls at list price).
+Methodology
**Dataset**
PWS-Legal is Parallel's private legal benchmark: 238 held-out questions with verified answers. 98 cover the companies, laws, regulations and IP rights legal researchers look up (60 about jurisdictions outside the US, 33 in 12 languages other than English), and 140 cover court decisions and legal documents on sites that general web search reaches poorly. The questions and answers are not published anywhere online.
**Evaluation methodology**
Every search API runs in the same agent loop with the same questions and the same LLM judge, with a budget of 20 tool calls per question, so the search layer is the only variable. Every arm answers with GPT-6 Astra (reasoning: high). Each API uses its own published tool definition at default settings, with 10 results per search; Perplexity is the Search API only. Firecrawl uses Firecrawl's own MCP search and scrape tools. Strict scoring: an answer is correct only if every required fact is right. Scores are the mean of three runs (Firecrawl is a single run).
Cost is end-to-end per answer at list price: the answer model's tokens plus every search and fetch the agent makes, shown per 1,000 questions. Firecrawl is priced at Standard-plan credits at one credit per scraped page; PDFs bill per page, so its real cost is higher.
**Testing dates**
Evals were run in November 2026.
**Dataset**
PWS-Legal is Parallel's private legal benchmark: 238 held-out questions with verified answers. 98 cover the companies, laws, regulations and IP rights legal researchers look up (60 about jurisdictions outside the US, 33 in 12 languages other than English), and 140 cover court decisions and legal documents on sites that general web search reaches poorly. The questions and answers are not published anywhere online.
**Evaluation methodology**
Every search API runs in the same agent loop with the same questions and the same LLM judge, with a budget of 20 tool calls per question, so the search layer is the only variable. Every arm answers with GPT-6 Astra (reasoning: high). Each API uses its own published tool definition at default settings, with 10 results per search; Perplexity is the Search API only. Firecrawl uses Firecrawl's own MCP search and scrape tools. Strict scoring: an answer is correct only if every required fact is right. Scores are the mean of three runs (Firecrawl is a single run).
Cost is end-to-end per answer at list price: the answer model's tokens plus every search and fetch the agent makes, shown per 1,000 questions. Firecrawl is priced at Standard-plan credits at one credit per scraped page; PDFs bill per page, so its real cost is higher.
**Testing dates**
Evals were run in November 2026.
### PWS-Legal (Parallel for Legal, Nov 2026)
| Series | Model | Cost (CPM) | Accuracy (%) | | -------- | ----------------- | ---------- | ------------ | | Parallel | Parallel Fast | 230.1 | 85.4 | | Parallel | Parallel Advanced | 194.2 | 83.1 | | Others | OpenAI | 440.8 | 80.5 | | Others | Firecrawl | 314.4 | 77.3 | | Others | Exa Auto | 318.9 | 75.6 | | Others | Tavily | 425.7 | 75.4 | | Others | Perplexity | 248 | 73.4 |
CPM: USD per 1,000 questions, end to end (answer-model tokens plus search and fetch calls at list price).
**Dataset**
PWS-Legal is Parallel's private legal benchmark: 238 held-out questions with verified answers. 98 cover the companies, laws, regulations and IP rights legal researchers look up (60 about jurisdictions outside the US, 33 in 12 languages other than English), and 140 cover court decisions and legal documents on sites that general web search reaches poorly. The questions and answers are not published anywhere online.
**Evaluation methodology**
Every search API runs in the same agent loop with the same questions and the same LLM judge, with a budget of 20 tool calls per question, so the search layer is the only variable. Every arm answers with GPT-6 Astra (reasoning: high). Each API uses its own published tool definition at default settings, with 10 results per search; Perplexity is the Search API only. Firecrawl uses Firecrawl's own MCP search and scrape tools. Strict scoring: an answer is correct only if every required fact is right. Scores are the mean of three runs (Firecrawl is a single run).
Cost is end-to-end per answer at list price: the answer model's tokens plus every search and fetch the agent makes, shown per 1,000 questions. Firecrawl is priced at Standard-plan credits at one credit per scraped page; PDFs bill per page, so its real cost is higher.
**Testing dates**
Evals were run in November 2026.
### Legal Hallucinations (Parallel for Legal, Nov 2026)
| Series | Model | Cost (CPM) | Accuracy (%) | | -------- | ----------------- | ---------- | ------------ | | Parallel | Parallel Fast | 20.01 | 96.5 | | Parallel | Parallel Advanced | 21.69 | 96.8 | | Others | Exa Auto | 28.12 | 96.8 | | Others | OpenAI | 43.79 | 97.5 | | Others | Perplexity | 20.11 | 94 | | Others | Firecrawl | 22.47 | 95.4 | | Others | Tavily | 86.64 | 90.1 |
CPM: USD per 1,000 questions, end to end (answer-model tokens plus search and fetch calls at list price).
**Dataset**
Stanford RegLab's Legal Hallucinations[Legal Hallucinations] dataset (Dahl et al., 2024) asks verifiable questions about real federal court cases, such as the citation, the court, the author, and whether the case exists. The original study found that LLMs answering without search hallucinated 58–88% of the time. Results are reported on a 284-question sample.
**Evaluation methodology**
Multi-step agentic evaluation: every search API runs in the same agent loop with the same questions and the same LLM judge, so the search layer is the only variable. The agent calls the provider's search tool plus its fetch tool as often as needed, and every arm answers with GPT-6 Sol (reasoning: high). Each API uses its own published tool definition at default settings, with 10 results per search; Exa runs in auto mode with an objective. Firecrawl uses Firecrawl's own MCP search tools, including its Government Index of US law. Results that copy the benchmark are removed before the model sees them; OpenAI native search can't be filtered this way, so its results were audited afterwards. Scores are from a single run.
Cost is end-to-end per answer at list price: the answer model's tokens plus every search and fetch the agent makes, shown per 1,000 questions. Firecrawl is priced at Standard-plan credits.
**Testing dates**
Evals were run in November 2026.
### LegalBench Citation (Parallel for Legal, Nov 2026)
| Series | Model | Cost (CPM) | Accuracy (%) | | -------- | ----------------- | ---------- | ------------ | | Parallel | Parallel Fast | 15.02 | 93.6 | | Parallel | Parallel Advanced | 16.54 | 93.8 | | Others | Perplexity | 16.52 | 87.2 | | Others | Exa Auto | 23.17 | 79.7 | | Others | Tavily | 27.92 | 74.7 | | Others | Firecrawl | 8.05 | 63.1 |
CPM: USD per 1,000 questions, end to end (answer-model tokens plus search and fetch calls at list price).
**Dataset**
LegalBench[LegalBench] (Guha et al., 2023) is a benchmark of 162 legal reasoning tasks designed by legal professionals at Stanford and other institutions. We ran the 161 questions in its citation tasks, which ask which case a federal appeals court cited for a legal proposition, and whether a given case supports it.
**Evaluation methodology**
One-search evaluation: every search API gets a single search call per question, then the model answers, with the same questions and the same LLM judge, so the search layer is the only variable. Every arm answers with GPT-6 Sol (reasoning: high). Each API uses its own published tool definition at default settings, with 10 results per search; Exa runs in auto mode with an objective. Firecrawl uses Firecrawl's own MCP search tools, including its Government Index of US law. Results that copy the benchmark are removed before the model sees them. Scores are the mean of three runs.
Cost is end-to-end per answer at list price: the answer model's tokens plus the search call, shown per 1,000 questions. Firecrawl is priced at Standard-plan credits.
**Testing dates**
Evals were run in November 2026.
### LegalBench: Finding the right case law with a single search
Legal AI needs to support its answers with the right case law. To test how well each search provider retrieves that evidence, we ran LegalBench[LegalBench], a benchmark of 162 legal reasoning tasks designed by legal professionals at Stanford and other institutions. We evaluated 161 questions from its citation tasks, which ask an agent to find the case that supports a legal statement.
### Legal Hallucinations: Grounding legal answers in verifiable facts
Finding the right case is only part of the challenge. Legal AI also needs to answer questions about that case accurately. Stanford RegLab built this benchmark[this benchmark] around verifiable questions about real federal court cases. The original study found that LLMs answering without search hallucinated 58–88% of the time. We evaluated 284 questions to test how well search helps agents get those facts right.
### PWS-Legal: Highest accuracy at up to 2.3× lower cost
Evaluating whether a model can handle a specific task is one of the hardest parts of building AI applications. Legal AI adds its own requirements for accurate search results. To help developers evaluate performance on legal tasks, we built PWS-Legal: 238 legal-specific questions, with 98 generated questions and 140 drawn from a specialized crawl of legal sources.
## How we reduce bias in the PWS-Legal benchmark
Vendor benchmarks tend to favor the vendor that builds them. To avoid that, we sourced every question from real research demand rather than Parallel's strengths, verified each answer through independent research routes, and ran every provider through the same harness with its published best practices, so only the search provider changes.
- - **Detailed questions.** 98 questions cover the companies, laws, regulations and IP rights legal researchers look up, across patents and IP, company records, regulation, legislation, court procedure and contracts.
- - **Global coverage.** 60 of those 98 questions are about jurisdictions outside the US, and 33 are in 12 languages other than English.
- - **Hard-to-reach sources.** 140 more questions cover court decisions and legal documents on sites that general web search reaches poorly.
- - **Verified answers.** Independent research routes, critics and a verifier build each answer. We drop any question that every search API already answers, Parallel included.
Here are five sample questions in the style of the set. They illustrate the question types; none are from the held-out set.
| Type | Question | Answer |
|---|---|---|
| Court procedure (US) | Under the Federal Rules of Appellate Procedure, how long does a party in a federal civil case have to file a notice of appeal? | 30 days after entry of the judgment or order; 60 days if the United States, a US agency, or a US officer or employee is a party (FRAP 4(a)(1)). |
| Patents and IP (US) | As of 3 November 2026, how long is the term of a US utility patent on an application filed today? | 20 years from the US filing date, or from the earliest non-provisional US or PCT application it claims. Term adjustment or extension can lengthen it; it ends early if maintenance fees lapse (35 U.S.C. 154). |
| Company records (US) | What is Apple Inc.'s SEC Central Index Key, and in which state is it incorporated? | CIK 0000320193; incorporated in California. |
| Regulation (EU) | As of 3 November 2026, when did the EU AI Act enter into force, and from when do its bans on prohibited AI practices apply? | In force 1 August 2024. The original bans apply from 2 February 2025; two bans added in 2026 apply from 2 December 2026. |
| Legislation (India) | Which law replaced the Indian Penal Code, and when did it take effect? | The Bharatiya Nyaya Sanhita, 2023, in force 1 July 2024 (except section 106(2)). |
## Legal responses depend on high quality sources
Legal teams need authoritative citations to verify search results and use them to inform decisions. A page that mentions the right words isn’t enough, and the agent needs the specific case or an opinion that cites it for the relevant legal rule. The two LegalBench examples below show this scenario in practice.
### Exa vs. Parallel Search on finding the case the Third Circuit cited
LegalBench question: Which case did the Third Circuit cite for "our role in interpreting an insurance policy is to 'ascertain the intent of the parties...'"? Correct answer: Am. Auto. Ins. Co. v. Murray. One search call per question, GPT-6 Sol answering, three runs each. Exa: 1 of 3 runs right. Its 8 results were all copies of the 2023 opinion that quotes the rule, so the model named that opinion instead of Murray. Parallel Search: 3 of 3 runs right. Its 10 results included Murray itself, two Third Circuit opinions citing Murray for this rule, and 7 other related documents, so the model could see Murray was the original source.
### Exa vs. Parallel Search on checking a quote against Murray
LegalBench question: Does Murray stand for "One element of the case-or-controversy requirement is that [plaintiffs]... must establish that they have standing to sue"? Correct answer: No. One search call per question, GPT-6 Sol answering, three runs each. Exa: 0 of 3 runs right. Its 9 results were mostly district-court opinions, none of them Murray, so the model had no copy of Murray to check the quote against. Parallel Search: 3 of 3 runs right. It returned Am. Auto. Ins. Co. v. Murray and Raines v. Byrd, the Supreme Court case the quote comes from, so the model could see the quote belongs to Raines, not Murray.
## How teams deploy Parallel for Legal in production
Agents built on Parallel already research case law across jurisdictions, answer legal questions from trusted legal and medical sources, and enrich case records for plaintiff firms. Here is how teams run Parallel in production today.
Harvey
1 of 3
AI platform for legal and professional services
Reaching legal sources across 60+ countries
Harvey uses Parallel Search and Extract to power web search across its platform, plus a specialized private index for hard-to-reach international sources, such as court rulings from Brazil's state tribunals and regulatory codes from Argentina.
“Court rulings, regulatory codes, and legislation in dozens of countries are published on sites that no search engine has ever indexed. Parallel is solving that for us at a scale we couldn't build in-house.” Gabe Pereyra, President & Co-Founder, Harvey
## Build your first jurisdiction survey
Our open-source cookbook shows how to use Parallel's APIs to compare one legal question across many jurisdictions. From a CSV of states and a question, such as the statute of limitations for a motor vehicle personal injury claim, the recipe runs Search with Source Policy limited to official legislature and court sites. The cookbook then takes these sources, and runs the Task API to produce a cited comparison table and a research basis page listing every source.
It's intended as a lightweight illustration of workflows we've seen work well in production. The code is in our open-source repo[our open-source repo].
## Get started
Start building for free[Start building for free] or dive into our docs[docs].
Contact us[Contact us] to learn more about enterprise plans with SOC 2 Type 2, zero data retention, a BAA for HIPAA, EU data residency, SSO, and specialized indexes for legal domains.
