Exa vs Parallel

We ran five public benchmarks between July 8 and July 24, 2026. The three below test multi-hop and hard factual retrieval. Each pairs an Exa configuration with the Parallel tier at a comparable latency budget, so speed stays roughly constant.

To ensure fairness, we use independent datasets. FRAMES was published by Google and SealQA by an outside academic team, and both are openly available, so anyone can rerun what is below and check the numbers themselves. We ran the evaluations. We did not design them.

Fast budget

Exa Fast returns in ~450ms, Parallel basic in ~900ms to 1.1s.

Research budget

Exa Auto returns in ~1.5s, Parallel advanced in ~2.2s.

Exa leads every one of these six comparisons while returning faster in both brackets. The margin runs from 1.4x to 2.5x, and it widens as the questions get harder.

Evaluation methodology:

• The public datasets scored here were built by outside teams. FRAMES comes from Google, SealQA and SealQA-Hard from Tu Vu's research group, BrowseComp from OpenAI, WideSearch from ByteDance Seed, DeepSearchQA from Google DeepMind, and Finance Agent v2 from Vals AI.
• Where no good public benchmark exists for a retrieval task, we build and open-source our own. Those live at github.com/exa-labs/benchmarks.
• Each API configuration was scored against the dataset's primary ground-truth answers.
• p50 latency and request cost were tracked concurrently with retrieval quality.
• Provider errors are counted and published in the results rather than dropped from the sample.
• Test suite run across eight providers: Exa, Parallel, Brave, Tavily, Perplexity, Linkup, Firecrawl, and You.com.

On the SimpleQA run, Exa and Parallel completed every query. Perplexity returned errors on 115 of 200, You.com on 95, and Firecrawl on 10.

Platform design reflects different architectural philosophies. Exa provides high-speed primitives for an agent to orchestrate. Parallel accepts a high-level objective and manages the execution loop autonomously.

Exa is the optimal choice for production agents where latency is a bottleneck, for complex research requiring semantic depth, and for vertical retrieval of people, companies, or code.

Exa for latency-critical agent loops

Exa Instant delivers a p50 latency of 306ms on SealQA. An agent performing a 50-search chain finishes its workflow in approximately 15 seconds.

Speed costs nothing in quality here:

• Exa Instant is 3x faster than Parallel basic while scoring 2.5x higher on accuracy.
• Exa Auto outperforms Parallel advanced on both speed and accuracy (2x higher score).
• Across the benchmarks above, no Parallel tier is both faster and more accurate than the Exa configuration nearest it.

Sub-350ms response times change what you can build. Voice agents can retrieve context without breaking conversational flow, and deep research agents can iterate through dozens of steps in seconds.

How Parallel compares: the same 50-call loop takes about 45 seconds on Parallel basic, and nearly two minutes on advanced.

Exa for people and company search

Exa maintains proprietary indexes for people and companies. A query like "Series B fintech companies in Singapore with 50-200 employees" retrieves structured records instead of keyword-matched web pages.

Index highlights:

• 1B+ profiles with 50M+ weekly updates.
• Company search powered by custom embeddings, filterable by headcount, funding, and industry.
• Full retrieval benchmarks available at github.com/exa-labs/benchmarks.

These are Exa's own benchmarks, built because no good public dataset covers entity retrieval, and open-sourced in full at github.com/exa-labs/benchmarks. Measured on 1,400 people queries and 605 company retrieval queries. The company dataset focuses on Series A/B startups and regional B2B players to test retrieval quality rather than model memory.

How Parallel compares: the core Search API has no people or company category. Entity data comes from separate products, Datasets and Enrichment, with Entity Search and FindAll in beta. Parallel reports FindAll results on its own 40-query set rather than on an open dataset.

Exa for multi-hop research questions

Search APIs are used for queries LLMs cannot answer from pre-training, such as linking an acquisition announcement to a specific figure in a subsequent earnings report.

Exa's neural retrieval handles these chains, and the gap widens as difficulty increases. All three comparisons below are Exa Auto against Parallel advanced, at a comparable latency budget:

• FRAMES (multi-source reasoning): 0.46 vs 0.30 (1.6x lead)
• SealQA: 0.27 vs 0.14 (2.0x lead)
• SealQA-Hard: 0.36 vs 0.19 (1.9x lead)

How Parallel compares: Parallel positions its Task API for multi-hop deep research, and prices it as a research product rather than a search call.

Exa for code retrieval in coding agents

Agents require specific function signatures rather than full documentation pages. Exa Code searches 1B+ pages and returns succinct excerpts based on a per-call token budget.

WebCode, published March 2026, roughly 840 queries across four retrieval tracks.

Exa leads on Signal by 17 points, which measures the density of relevant information per token. That is the number that matters when an agent pays per token. Grounding with Exa Code increased accuracy to 73% across a range of programming libraries.

import requests

response = requests.post(
    "https://api.exa.ai/context",
    headers={"x-api-key": "your-api-key"},
    json={
        "query": "React Server Components with Next.js App Router",
        "tokensNum": 5000
    }
)
print(response.json()["response"])

How Parallel compares: Parallel has no code-specific search endpoint, though accuracy ties at 89 and Parallel leads table recall by 0.3.

Exa Agent for research on a fixed budget

The Exa Agent ingests high-level research objectives and an effort parameter, ranging from minimal to xhigh, to deliver comprehensive, autonomous research findings.

With a cost of just $0.50 per request, the High configuration either equals or surpasses the retrieval quality of Parallel's premium tiers across all evaluated benchmarks.

Published in Introducing Exa Agent, June 2026.

On BrowseComp (1,266 queries), Exa Agent High scores 74% for $0.50, while Parallel Task Ultra 8x scores 58% for $2.40. Exa is 16 points more accurate at roughly a fifth of the cost.

How Parallel compares: Parallel reaches its best score here on Ultra 8x, the top processor tier, which is built for tasks that justify far more compute per request.

Exa for academic and publication search

Exa maintains a proprietary index of 350M+ academic works, prioritizing direct retrieval of original papers over secondary summaries. This performance was validated across 1,472 queries designed to test both high-precision known-item identification and "tip-of-the-tongue" retrieval from imprecise or partially erroneous descriptions.

Exa achieves 36 points higher recall than Parallel advanced and operates 5x faster.

How Parallel compares: Turbo is the only configuration faster than Exa in this set, but it retrieves the correct paper only 39.2% of the time.

Parallel is a robust platform with two distinct areas of strength.

A higher compute ceiling per task

Parallel's Task API escalates through named processor tiers (Core to Ultra 8x), so a single task can be given far more compute than a fixed effort level allows. That fits workloads where a small share of tasks deserve much more work than the rest.

How Exa compares: Exa Agent uses predictable effort levels that top out at xhigh.

Premium extraction from difficult pages

Parallel's Extract API is particularly effective with PDFs and heavy JavaScript documentation.

Other platform features include:

• Deep research Task API with provenance tracking.
• Monitor API for web change detection.
• Dataset enrichment for structured table building.

How Exa compares: Exa's Contents API handles PDFs and JavaScript-rendered pages automatically as part of search, rather than as a separate product to call.

A unified semantic index powering four distinct API products.

Exa Search API

Exa Search provides the highest accuracy of any search API within its latency class. It processes natural language queries to deliver results optimized for token efficiency, complete with citations.

Users can select from six search types spanning budgets from ~200ms to 60 seconds. This includes instant and fast for speed-sensitive loops, auto for adaptive scaling, and a range of deep research modes: deep-lite, deep, and deep-reasoning.

Exa Agent API

The Exa Agent is a high-compute research primitive designed to synthesize information from the entire web. It automates multi-hop reasoning for complex workflows like exhaustive list-building and entity enrichment.

Compute is governed by effort levels ranging from minimal to xhigh, alongside an auto mode that scales resources to meet the task requirements.

Exa Contents API

The Contents API retrieves full-page text, summaries, and structured data according to a user-defined schema. It automatically resolves complex layouts, including PDFs and pages requiring JavaScript rendering.

Query-dependent highlights reduce token consumption significantly; just 500 characters of highlights provide the same retrieval accuracy as the initial 8,000 characters of page content, all with sub-100ms latency.

Exa Monitors API

The Monitors API provides real-time visibility into web changes, surfacing updates the moment they occur.

Teams building production-ready agents requiring high-speed, diverse search content will typically find Exa to be the stronger alternative. Parallel is well-suited for users who prefer managed research loops with per-task compute scaling.

Migration is mostly a mapping exercise. The two APIs take different request shapes: Parallel takes an objective with search_queries and nested advanced_settings, while Exa takes a top-level query with its parameters alongside it.

You'll want to:

• Find equivalent parameters in the Exa API. Note that Exa's API products tend to be more configurable than Parallel's API schema. A few notes:
  • Parameters have different names. Exa uses numResults, includeDomains, and category.
  • Date filtering uses start/end dates instead of after_date.
  • Addition of includeText/excludeText and language filters.
• Swap out the Parallel "processor" tier:
  • If switching from the Parallel Task API or FindAll API, replace the processor tier with an Exa Agent effort level. For most tasks, more affordable Exa Agent effort levels outperform the closest equivalent processor tier.
  • If switching from the Parallel Search API or Responses API, choose a search type from the Exa Search API that matches your latency budget. We recommend selecting auto by default (which provides reliable results at ~1s latency).

Integrations with LiteLLM, Vercel AI SDK, LangChain, and many other providers are natively supported, simplifying the cutover process.