# Upgrades to the Parallel Search & Extract APIs

Tags:Benchmarks
Reading time: 5 min
Search & Extract Benchmarks

Today, we’re announcing major upgrades to the quality and capabilities of our Search and Extract APIs, and making both generally available. These enhancements deliver industry-leading accuracy and token efficiency for web search across diverse use cases, including company research, coding agents, multilingual queries, finance, and broad internet research.

## Highlights

  • - New streamlined search modes catering to interactive and background agents
  • - Pareto optimal performance on challenging public benchmarks
  • - Leading performance on benchmarks focused on specialized knowledge work
  • - Superior compression of webpage context using the Extract API
  • - Improvements to global index coverage and multi-lingual capabilities

## The best web search for AI agents in interactive and background use cases

As AI agents have become better at handling knowledge work that needs the web, two main product patterns have emerged: interactive workflows and long-running background tasks.

Interactive workflows support real-time collaboration between a user and an agent. They are shaped by latency requirements and frequently focus on targeted web research or fetching specific information.

Background tasks work more independently on ambiguous, multi-step questions spanning several topics. These searches typically require deeper retrieval and orchestration of multiple tool calls.

The Search API now offers two modes designed for these distinct use cases.

**Basic mode** is optimized for quick retrieval. It delivers P50 latency under one second and P90 under two seconds, making it a good fit for interactive, human-in-the-loop experiences.

**Advanced mode** is built for depth. It spends more time querying, reranking, and compressing results across both general-purpose and specialized indexes. Traditional search APIs often force agents into a series of sequential searches, which increases latency and expands the context window. Advanced mode resolves more in a single call.

Search ModeMedian LatencyBest for
Basic1sSingle-hop, latency-sensitive agents, e.g., human-in-loop chat, interactive experiences
Advanced3sMulti-hop background agents that can tolerate extra latency for better depth, e.g., code review agents, deep research
BrowsecompHumanity's Last ExamSealQA-Seal0WebWalkerFreshQA

CPM: USD per 1000 requests. Cost is shown on a Log scale.

Others
Parallel
Benchmark comparison across Cost (CPM) and Accuracy (%). CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

We evaluated search providers against five open benchmarks covering complementary aspects of agentic search: BrowseComp (hard multi-hop questions that require navigating the live web), Frames (multi-document factoid reasoning), FreshQA (time-sensitive questions where the correct answer depends on recent web information), HLE (Humanity's Last Exam — expert-level academic questions spanning math, science, and humanities), SealQA (ambiguity-robust factoid QA with intentionally misleading snippets), WebWalker (tasks designed around following links across pages to find an answer).

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to `MAX_TOOL_CALLS=25` tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.


**Testing dates**

April 19-21, 2026

### Browsecomp

| Series   | Model             | Cost (CPM) | Accuracy (%) |
| -------- | ----------------- | ---------- | ------------ |
| Others   | Tavily            | 973        | 42           |
| Others   | Exa               | 1160       | 40           |
| Parallel | Parallel Basic    | 600        | 53           |
| Parallel | Parallel Advanced | 379        | 51           |

CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

We evaluated search providers against five open benchmarks covering complementary aspects of agentic search: BrowseComp (hard multi-hop questions that require navigating the live web), Frames (multi-document factoid reasoning), FreshQA (time-sensitive questions where the correct answer depends on recent web information), HLE (Humanity's Last Exam — expert-level academic questions spanning math, science, and humanities), SealQA (ambiguity-robust factoid QA with intentionally misleading snippets), WebWalker (tasks designed around following links across pages to find an answer).

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to `MAX_TOOL_CALLS=25` tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.


**Testing dates**

April 19-21, 2026

### Humanity's Last Exam

| Series   | Model             | Cost (CPM) | Accuracy (%) |
| -------- | ----------------- | ---------- | ------------ |
| Parallel | Parallel Basic    | 451        | 58           |
| Parallel | Parallel Advanced | 315        | 56           |
| Others   | Exa               | 522        | 57           |
| Others   | Tavily            | 538        | 54           |

CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

We evaluated search providers against five open benchmarks covering complementary aspects of agentic search: BrowseComp (hard multi-hop questions that require navigating the live web), Frames (multi-document factoid reasoning), FreshQA (time-sensitive questions where the correct answer depends on recent web information), HLE (Humanity's Last Exam — expert-level academic questions spanning math, science, and humanities), SealQA (ambiguity-robust factoid QA with intentionally misleading snippets), WebWalker (tasks designed around following links across pages to find an answer).

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to `MAX_TOOL_CALLS=25` tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.


**Testing dates**

April 19-21, 2026

### SealQA

| Series   | Model             | Cost (CPM) | Accuracy (%) |
| -------- | ----------------- | ---------- | ------------ |
| Parallel | Parallel Basic    | 258        | 45           |
| Parallel | Parallel Advanced | 191        | 41           |
| Others   | Tavily            | 243        | 45           |
| Others   | Exa               | 326        | 41           |

CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

We evaluated search providers against five open benchmarks covering complementary aspects of agentic search: BrowseComp (hard multi-hop questions that require navigating the live web), Frames (multi-document factoid reasoning), FreshQA (time-sensitive questions where the correct answer depends on recent web information), HLE (Humanity's Last Exam — expert-level academic questions spanning math, science, and humanities), SealQA (ambiguity-robust factoid QA with intentionally misleading snippets), WebWalker (tasks designed around following links across pages to find an answer).

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to MAX_TOOL_CALLS=25 tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.

**Testing dates**

April 19-21, 2026

### WebWalker

| Series   | Model             | Cost (CPM) | Accuracy (%) |
| -------- | ----------------- | ---------- | ------------ |
| Others   | Exa               | 210        | 74           |
| Others   | Tavily            | 202        | 71           |
| Parallel | Parallel Advanced | 101        | 73           |
| Parallel | Parallel Basic    | 155        | 71           |

CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

We evaluated search providers against five open benchmarks covering complementary aspects of agentic search: BrowseComp (hard multi-hop questions that require navigating the live web), Frames (multi-document factoid reasoning), FreshQA (time-sensitive questions where the correct answer depends on recent web information), HLE (Humanity's Last Exam — expert-level academic questions spanning math, science, and humanities), SealQA (ambiguity-robust factoid QA with intentionally misleading snippets), WebWalker (tasks designed around following links across pages to find an answer).

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to `MAX_TOOL_CALLS=25` tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.


**Testing dates**

April 19-21, 2026

### FreshQA

| Series   | Model             | Cost (CPM) | Accuracy (%) |
| -------- | ----------------- | ---------- | ------------ |
| Parallel | Parallel Advanced | 49         | 79           |
| Parallel | Parallel Basic    | 90         | 77           |
| Others   | Exa               | 84         | 78           |
| Others   | Tavily            | 89         | 78           |

CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

We evaluated search providers against five open benchmarks covering complementary aspects of agentic search: BrowseComp (hard multi-hop questions that require navigating the live web), Frames (multi-document factoid reasoning), FreshQA (time-sensitive questions where the correct answer depends on recent web information), HLE (Humanity's Last Exam — expert-level academic questions spanning math, science, and humanities), SealQA (ambiguity-robust factoid QA with intentionally misleading snippets), WebWalker (tasks designed around following links across pages to find an answer).

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to `MAX_TOOL_CALLS=25` tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.


**Testing dates**

April 19-21, 2026

### FRAMES

| Series   | Model             | Cost (CPM) | Accuracy (%) |
| -------- | ----------------- | ---------- | ------------ |
| Parallel | Parallel Advanced | 93         | 87           |
| Parallel | Parallel Basic    | 165        | 84           |
| Others   | Exa               | 169        | 87           |
| Others   | Tavily            | 189        | 83           |

CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

We evaluated search providers against five open benchmarks covering complementary aspects of agentic search: BrowseComp (hard multi-hop questions that require navigating the live web), Frames (multi-document factoid reasoning), FreshQA (time-sensitive questions where the correct answer depends on recent web information), HLE (Humanity's Last Exam — expert-level academic questions spanning math, science, and humanities), SealQA (ambiguity-robust factoid QA with intentionally misleading snippets), WebWalker (tasks designed around following links across pages to find an answer).

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to `MAX_TOOL_CALLS=25` tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.


**Testing dates**

April 19-21, 2026

## The best web search API for specialized knowledge work

We believe the best web search API for agents is a strong general-purpose system with the scale, depth, and ranking quality to perform well across many kinds of knowledge work.

Our API is built on that foundation. It offers broad coverage of the web while going deep in the domains that matter most, so agents can retrieve high-quality results for both general research and specialized tasks. Whether the job is debugging against current library documentation or researching a prospect’s funding history and tech stack, the same strengths apply: a large index, deep retrieval, and strong ranking.

Advanced mode is designed for the most demanding workflows, where agents benefit from deeper retrieval and more precise ranking across both broad web content and domain-relevant sources.

On our internal dataset focused on company search, Parallel’s Search API outperforms alternatives in finding highly relevant candidates that meet the specified search criteria.

Illustration demonstrating deep research API concepts, web search capabilities, or AI agent integration features

Company Benchmark Methodology

**Dataset**

Queries targeting companies across categories, filtering by sector, funding stage, geography, investors, revenue, and team composition. Queries were generated using an LLM.

**Sample queries**

The benchmark queries are intentionally complex — the kind of multi-constraint searches that break general-purpose web search:

  • - "Biotech and solar energy startups serving healthcare providers, backed by IVP"
  • - "Series B SaaS companies focused on manufacturers and industrial production."

**Evaluation setup:**

250 queries, 10 results per query, single-graded by GPT-5.4-mini. The LLM judge graded every returned URL on whether it represents a genuine match for the query objective.

We measure **precision** which represents the fraction of returned results that are relevant — the metric that matters most for candidate generation, where downstream workflows depend on signal quality.

Similarly, for queries generated by coding agents as part of development workflows, Parallel’s Search API provides the most relevant code snippets and documents.

Coding

CPM: USD per 1000 requests. Cost is shown on a Log scale.

Parallel
Others
Benchmark comparison across Cost (CPM) and Accuracy (%). CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

A proprietary coding dataset derived from production queries to Parallel’s search API.

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to `MAX_TOOL_CALLS=25` tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report the accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.

**Testing dates**

April 19-21, 2026

### Coding

| Series   | Model             | Cost (CPM) | Accuracy (%) |
| -------- | ----------------- | ---------- | ------------ |
| Parallel | Parallel Standard | 154        | 82           |
| Parallel | Parallel Basic    | 269        | 81           |
| Others   | Exa               | 331        | 80           |
| Others   | Tavily            | 352        | 75           |

CPM: USD per 1000 requests. Cost is shown on a Log scale.

**Dataset**

A proprietary coding dataset derived from production queries to Parallel’s search API.

**Evaluation methodology**

Every task is run through a shared deep-research harness: a single GPT-5.4 agent is given two tools (web search and web fetch) with an iterative budget of up to `MAX_TOOL_CALLS=25` tool calls per question. The agent plans sub-queries, fans out searches, fetches specific pages when snippets are insufficient, and returns an answer when it exhausts the number of allowed tool calls or has sufficient information to answer the question. Each answer is then LLM-graded by GPT-5.4. We report the accuracy of the final answer.

We measure accuracy and overall cost, which includes LLM token costs and tool call costs.

**Testing dates**

April 19-21, 2026

## Combine Search with Extract for the deepest, most token-efficient multi-hop agent

Search and Extract are designed to work together. Search finds the right sources with high precision. Extract turns those sources into concise, task-relevant context that an agent can use efficiently.

The Parallel Extract API is especially useful when the source is long or dense, such as a 200-page SEC filing, an investor document, or a complex technical page. Passing the full source into context increases token usage and makes downstream reasoning less focused.

The Parallel Extract API solves this by letting the agent specify an extraction objective. After Search identifies the right source, Extract returns only the content most relevant to that objective. The result is a more efficient retrieval pipeline: Search provides breadth and precision, while Extract compresses large pages and documents into high-signal context for the next reasoning step.

Together, they give agents the best of both: high-quality retrieval and high-quality compression in a single workflow.

On FinanceBench[FinanceBench](https://docs.patronus.ai/docs/research_and_differentiators/financebench), built around real-world workflows across SEC filings, earnings reports, and multi-page investor documents, the extract API consistently succeeds at getting the highest recall for any number of tokens returned.

Illustration demonstrating deep research API concepts, web search capabilities, or AI agent integration features

Finance Benchmark Methodology

**Dataset**

The eval sampled 136 questions from the finance bench QA dataset.

 

**Evaluation setup**

For each evaluation question, the harness runs the extract APIs against the question's URLs and objective, then judges the response using an LLM-as-judge based on the reference answer.

 

**Evaluation**

1. Combine all extracted content returned by the engine into a single excerpt.

2. LLM judge (GPT 5.4 mini) determines whether the combined excerpt contains enough information to correctly answer the question, given a known reference answer. The judge returns a short chain of reasoning plus a binary verdict: *sufficient* or *insufficient*.

3. Aggregate recall across the dataset is the fraction of questions the extracted content could answer. If the engine returns no content, the question is scored zero without calling the judge.

## We’ve expanded Parallel Search to more corners of the world

Sophisticated AI products serve users globally. A GTM agent may prospect across markets, a compliance agent may track filings across jurisdictions, and a research assistant may serve users in many languages. These workflows require a search layer that can retrieve fresh, relevant content across languages and regions with consistent quality.

The Parallel Search API now supports this natively, giving agents a stronger foundation for global workflows.

  • - **Global index coverage: **Our index now spans web pages from 30+ countries, including Korean, Chinese, Japanese, Spanish, Indian, German, Arabic, Portuguese, and Russian language content.
  • - **Multilingual query input: **Submit queries in any language. The Search API handles retrieval across languages without additional configuration.
  • - **Location parameter: **Set a target location to refine results to a specific region.
Illustration demonstrating deep research API concepts, web search capabilities, or AI agent integration features

Multi-Lingual Benchmark Methodology

**Dataset**

Starting from OpenAI's **SimpleQA**, each question was translated into all 26 languages from Apple's **MKQA** benchmark. Translations were carried out using LLMs and were spot checked by human annotators fluent in the respective languages. The reference answer is also translated/localized per language, giving us paired queries where the _same fact_ is asked across Latin-European, Cyrillic, CJK, Semitic (RTL), and SE Asian scripts.

 

**Evaluation methodology**

100 questions from each language; each query is sent to the search provider as-is, and the results are fed into an LLM for synthesis. The final answer is graded using an LLM-as-a-judge.

 

We report 2 metrics:

  • - **The fraction of answers marked as correct**, as reported by the judge. The grader sees the question, the gold answer, and the predicted answer, and returns CORRECT / INCORRECT / NOT_ATTEMPTED. We grade with an LLM rather than string EM/F1 because the agent's predicted answers are full sentences in many scripts — "The 2010 recipient was Michio Sugeno" in English, ميتشيو سوغينو transliterated variants in Arabic, etc. What we care about is whether the agent got the answer right, which the LLM judge captures directly.
  • - **Cross-lingual retrieval ability**: Because every question is paired across all 26 languages, the drop in accuracy when switching from English to non-English languages for the same question isolates the provider's cross-lingual retrieval quality. A small drop means the index and retrieval pipeline handle non-English queries nearly as well as English; a large drop means the provider is English-centric and falls off when asked to cross script families.

Parallel builds web infrastructure for AI. We give AI agents structured, grounded access to the open web, enabling agents to find, extract, monitor, and reason over information at a scale and quality no human workflow can match. Our infrastructure is powered by a rapidly growing proprietary index of the global internet, built from the ground up for AI consumption.

Fortune 100s and leading frontier AI companies, including Harvey, Manus, Modal, Starbridge, and Profound, rely on Parallel for grounding, fact-checking, contract monitoring, and high-quality content generation.

Get started on our Developer Platform[Developer Platform](https://platform.parallel.ai/) or read the documentation[documentation](https://docs.parallel.ai/).

Parallel avatar

By Parallel

April 21, 2026

## Related Posts60