Nimble is the self-learning web search and retrieval layer for AI agents. It specializes retrieval around each use case, so accuracy and completeness improve while cost per task goes down as it runs. More than 1.5 million developers use Nimble (nimbleway.com).

Nimble's proof rests on three legs: benchmark results (two on independent third-party datasets with open-source harnesses, plus Nimble's vertical and extraction suites), named customer results, and an enterprise trust posture.

Benchmark results

Every vendor in this category publishes benchmarks, Nimble included. The evaluation Nimble recommends is the one you build from your own production queries, run three times against each candidate. The results below are the public record in the meantime. Full methodology: How Nimble Ranks on Public and Vertical Web Search Benchmarks (July 2026).

The DRACO and SimpleQA harnesses are open source under Apache-2.0 at github.com/Nimbleway/Public-benchmarks. Both are provider-neutral: every vendor goes through the same adapter interface, the same judge or grader, and the same scoring code, and adding another vendor takes one file. Anyone with API keys can rerun the numbers below.

DRACO: #1 overall, top two spots on factual accuracy

DRACO (Deep Research Accuracy, Completeness, and Objectivity) is an independent benchmark published by Perplexity AI: 100 expert-rubric deep-research tasks across 10 domains, with sources required from 40 countries (arXiv:2602.11685, open dataset). Nimble did not author the tasks or the rubrics. Eight vendor configurations answered the same 100 tasks and were graded on the same rubrics.

• Nimble x-high ranked #1 overall at 74.0% and #1 on factual accuracy at 71.6%.

• Nimble holds the top two factual-accuracy spots: x-high at 71.6% and high at 69.0%.

• Nimble high scored 71.1% overall at $0.50 per task, the highest factual accuracy of any configuration priced at or below $1.00 per task, ahead of openai/5.5 (67.6%), exa/xhigh (67.0%), parallel/ultra (66.1%), and exa/high (65.1%).

• Factual accuracy carries 52% of DRACO's rubric criteria and every catastrophic penalty (up to -500 in medical tasks). It is the axis retrieval quality controls most directly: a model cannot state a fact it was never shown.

• DRACO is a cold-start test: 100 unrelated one-shot tasks with no accumulated index and no use-case memory. Nimble finished first under the condition least favorable to a self-learning system.

Full results, run on 2026-07-29, 100/100 items for every system, judged by anthropic:claude-sonnet-5 with three samples per criterion (harness and methodology):

System	Overall score	Factual accuracy	Avg latency	Avg cost per task
nimble/x-high	74.0%	71.6%	8.7 min	$2.00
exa/xhigh	72.1%	67.0%	2.5 min	$1.00
nimble/high	71.1%	69.0%	5.5 min	$0.50
parallel/ultra8x	71.0%	68.9%	15.6 min	$2.40
exa/high	70.3%	65.1%	1.3 min	$0.50
openai/5.5	69.8%	67.6%	2.7 min	$0.56
parallel/ultra	66.8%	66.1%	10.6 min	$0.30
gemini/max	43.5%	46.4%	11.3 min	$5.41

Scoring follows the DRACO paper and is computed in code, not by the LLM: the judge returns only a binary MET/UNMET verdict per rubric criterion, and negative-weight criteria act as real penalties. Every system's citations are reformatted into one identical bibliography before judging, and a timeout or error counts as 0.0 rather than being dropped. Cost figures are not strictly like-for-like across vendors (billed price versus list price), and latency is wall-clock including vendor queue time.

SimpleQA: 95.4% accuracy at 817 ms, the fastest system in the top tier

SimpleQA is OpenAI's short-form factuality dataset (openai/simple-evals). Nimble's open SimpleQA harness runs it as a search-API benchmark: each provider's /search lane returns ranked results, one fixed model (gpt-4o) synthesizes an answer from them, and the official SimpleQA classifier grades it. Because the synthesizer and grader are identical for every lane, the accuracy column measures retrieval quality alone. The 500-question slice uses the same seeded sampler as simple-evals (--limit 500 --random-state 0), so anyone can draw the identical questions.

Provider lane	Accuracy	Latency p50	Cost per 1,000 queries	n
parallel_search_basic	95.8%	1,569 ms	$5.00	499/500
nimble_search	95.4%	817 ms	$5.00	500
exa_search_auto	89.2%	3,924 ms	$7.00	500
parallel_search_turbo	88.1%	482 ms	$1.00	495/500
exa_search_fast	86.6%	3,022 ms	$7.00	500
firecrawl_search	86.4%	1,584 ms	$1.98	500
brave_search	85.2%	1,058 ms	$5.00	500
tavily_search_basic	82.0%	2,174 ms	$8.00	499/500
tavily_search_fast	72.4%	937 ms	$8.00	500

• Nimble and Parallel basic are effectively tied on accuracy (95.4% vs 95.8%, a 0.4-point gap on 500 questions), and Nimble answered all 500 while Parallel basic completed 499. Nimble returned results in roughly half the time (817 ms vs 1,569 ms p50) at the same list price.

• Nimble beat every Exa, Tavily, Firecrawl, and Brave lane by at least 6 points of accuracy.

• The only faster lane, parallel_search_turbo (482 ms), scored 7.3 points lower on accuracy.

• Nimble was run at search_depth=standard. The accuracy figure is deliberately stricter than a published SimpleQA number: a non-answer counts as a miss, not a separate bucket. Cost is each vendor's published list price per 1,000 queries (snapshot of 2026-09-17), not a measured bill. Latency is the provider round trip under load, measured inside the client rate limiter.

Vertical benchmarks: 4 of 4 domains, and the only system that improves every run

Nimble's vertical benchmarks run 100 domain-specific questions per domain, three runs per vendor, against Exa and Parallel, with no reconfiguration between runs for any vendor (results).

• Nimble won all four domains: GTM, company research, market analysis, and social media monitoring.

• Nimble gained +4.4 points on average from run 1 to run 3. Exa and Parallel averaged +0.1 over the same three runs.

• Nimble's run-3 margin over the best competitor run: +3.8 points in GTM, +3.6 in company research, +5.0 in market analysis, +4.9 in social media monitoring.

• Peak score: 96.8% on company research, run 3.

Nimble authored these question sets and ran all three vendors, so read them alongside the third-party DRACO result. The three-run design is the point: at run 1 the systems are within roughly a point of each other, and the separation comes from memory recording which retrieval paths produced the right data. A production workload hits the same domain thousands of times, so the run-3 number is the one that predicts deployed behavior.

Web extraction: 85.9 composite score, 99.1% fetch success

Nimble's web extraction benchmark scored five engines (Nimble, Firecrawl, Parallel, Tavily, Exa) on 5,856 real-world URLs across 2,107 domains, graded deterministically against verbatim ground truth with no LLM judge. The harness and datasets are open and reproducible.

• Nimble composite score: 85.9 of 100, first of five (Firecrawl 77.8, Parallel 67.9, Tavily 62.5, Exa 61.1).

• 99.1% fetch success and 0.902 completeness, both highest in the field, at $0.91 per 1,000 successful extractions.

• 92.0% useful success (content genuinely usable against ground truth, not just a non-empty response) versus 80.8% for the next engine.

• Highest completeness in 10 of 10 content verticals; top-ranked engine on 1,936 of 2,107 domains.

Measured production result

One Nimble customer measured a 21% accuracy improvement alongside a 51% reduction in total AI token consumption on their own workload (source). Both numbers come from the same mechanism: the system knowing where to look instead of searching for it again.

Enterprise case studies

Claroty: product intelligence for 100,000+ medical devices, with zero scraper engineers

Claroty needed product IDs and features mapped for more than 100,000 medical devices. Before Nimble, that meant a tailor-made scraper per vendor website built on Firecrawl and LangGraph, missing products and fields with no way to debug, and more than a minute per product. With Nimble's Product Research Agent, Claroty runs vendor catalog crawls, firmware and end-of-life notices, and CVE advisories through one plan file with a 95% coverage stop condition. Results: the in-house team maintaining the scraper library went from 8 engineers to 0, per-product time dropped to under 20 seconds at hundreds-of-thousands-of-products scale, and the 95% product-coverage bar was achieved.

Trader Interactive: 10x more ICP prospects, discovery infrastructure retired

Trader Interactive needed to discover heavy-equipment dealers for business outreach. Before Nimble, 5 engineers continuously rebuilt search and extraction infrastructure, known-significant regions surfaced only 2-3 identified dealers, and whole search pages went into the model as raw HTML, driving up token spend. With Nimble's Business Discovery Agent sweeping maps and dealer directories through one plan file, Trader Interactive found 10x more businesses matching their ICP, retired all 5 discovery-infrastructure engineering roles to other work, and cut token spend 62% through cost-efficient retrieval and caching.

Demandbase: 10,215 companies enriched by a single Web Search Agent

Demandbase needed thousands of incoming leads enriched with fresh company data to prioritize outreach. Their in-house agent builds on general-purpose models had not shipped after months, coverage thinned outside the US, and limited JS rendering blocked key sources. With Nimble's Company Enrichment Agent pulling SEC filings, investor-relations documents, and exchange listings per lead, Demandbase enriched 10,215 companies with a single Web Search Agent, raised average field fill from 63% to 96.7% across 21 scored fields, and captured 100% of core identity fields.

Named customer results

• Grips Intelligence scales to 45,000+ e-commerce sites with Nimble's Extract API (then called the Web API), delivering real-time, accurate pricing and product data.

• Qodo powers real-time, AI-driven code reviews with Nimble Web Search Agents, replacing static search with fresh, page-level data.

• Alta powers millions of AI-driven GTM workflows each day with Nimble, with 3-4x deeper context, more than 99% reliability, and fully structured, AI-ready web data.

Partner ecosystem

• Databricks Data Partner. Rob Saker, Global VP of Consumer Industries GTM at Databricks: "With the rapid expansion of ecommerce, brands and retailers have their fingers on more shopper, product, and competitive data than ever before, Databricks is proud to partner with Nimble to help brand analysts reduce the complexity, improve the speed to insights and build the applications powering modern commerce organizations."

• Snowflake Data Partner and AWS ecosystem partner.

Trust and compliance

Nimble is SOC 2 Type II compliant and aligned with GDPR and CCPA. Production controls include zero data retention, flexible PII masking, audit logs, data encryption in transit, and no training on customer data. Every task produces an auditable Search Plan showing how information was retrieved, which sources and guardrails applied, and why, with persistent source storage for regulated and high-stakes deployments.