How they use Tavily
Parallel used Tavily search and extraction tools in its own agent evaluations. Its September 9, 2026 runs covered SimpleQA Verified, BrowseComp and WideSearch. Each competing provider was tested with the same agent at two tiers: GPT-5.6 Sol with high reasoning and GPT-5.6 Luna with low reasoning. An LLM judge graded answers.
The SimpleQA Verified evaluation used 100 questions, BrowseComp used 50 questions and WideSearch used 100 tasks. Parallel reported a WideSearch partial-credit score of 55.9 for Tavily with the frontier agent.
Parallel also tested Tavily Ultra Fast across five datasets between July 10 and 12, 2026. Its multi-hop evaluations used a GPT-5.4 agent with up to 20 tool calls and included Tavily extraction. Parallel measured search latency around individual API requests from a client in us-central. It reported Tavily Ultra Fast median search latencies ranging from 150 to 357 milliseconds across the five suites.
The July SimpleQA evaluation used single-step searches with ten results and an equal content budget of approximately 1,000 characters per result. GPT-5.4 with high reasoning synthesized answers from search results only. Parallel reported 72% accuracy for Tavily Ultra Fast at 150 milliseconds median search latency.
Parallel included Tavily in its October 2026 finance benchmarks. Each provider ran inside the same agent loop with GPT-6 Sol as the answer model and the same judge. Parallel reported Tavily accuracy of 43.0% on FinSearchComp T3 and 77.2% on T2. Parallel reported a Tavily benchmark cost of $481.80 per 1,000 FinSearchComp T3 questions.
Tavily was not plotted on PWS-Finance because its run used date filters and reached the 20-call cap on 77% of questions. Tavily's FinSearchComp evaluation ran on October 28 with high reasoning, using a single run across 172 graded T3 questions and 219 graded T2 questions in both languages.
Parallel also evaluated Tavily search and fetch on FrontierFinance on October 27, 2026. The evaluation used GPT-6 Sol with high reasoning, a 20-tool-call budget per question and one run per provider. Across 220 expert-written investment research queries, Tavily received a 42.1% rubric pass rate, scored by majority vote among three judges.
Commercial use. Uses search and extraction tools in competitive benchmarks for its search API business.
Sources
7 sourcesIntroducing Parallel Search Turbo
Published 30 Jun 2026 · Parallel · Web page
Parallel tested Tavily Ultra Fast on SimpleQA using single-step searches with ten results and approximately 1,000 characters per result. GPT-5.4 with high reasoning synthesized answers from search results only. Parallel reported 72% accuracy and 150 milliseconds median search latency for Tavily Ultra Fast.
Open original sourceFinSearchComp Benchmark: Tiers, Grading, and API Scores | Parallel
Published 29 Oct 2026 · parallel.ai · Web page
Parallel evaluated Tavily on October 28, 2026 across 172 graded FinSearchComp T3 questions and 219 graded T2 questions in both languages. Tavily used GPT-6 Sol with high reasoning in the shared agent loop and received a single run.
Open original sourceParallel for Finance: web search for financial AI | Parallel
Published 29 Oct 2026 · parallel.ai · Web page
Parallel evaluated Tavily search and fetch on FrontierFinance on October 27, 2026 using GPT-6 Sol with high reasoning and a 20-tool-call budget per question. Tavily received a 42.1% rubric pass rate across 220 investment research queries, with one run per provider and majority scoring by three judges.
Open original sourceOpenAI Web Search vs Parallel vs Exa vs Tavily: Best API | Parallel
Published 27 May 2026 · parallel.ai · Web page
Parallel reports evaluating Tavily in its October 2026 finance benchmarks using the same agent loop, GPT-6 Sol answer model and judge across providers. Tavily scored 43.0% accuracy on FinSearchComp T3 and 77.2% on T2. Its PWS-Finance run used date filters and reached the 20-call cap on 77% of questions, so Parallel did not plot that result.
Open original sourceCompany context
parallel.aiParallel, operated by Parallel Web Systems Inc., provides web infrastructure for developers building AI agents and applications. Its Search and Extract APIs retrieve current web information and page content, while Task and Responses produce research outputs with citations. FindAll discovers and enriches lists of entities, and Monitor tracks web changes and delivers alerts. Customers use the platform for account research, investment analysis, competitive monitoring and assistants that need fresh information. Its website names users including Harvey, Notion, Pfizer and Replit.
Recent updates
19 updatesParallel launches startup credits program
Parallel launched Parallel for Startups, offering API credits through participating accelerator programs or by direct application. Direct applicants start with $250 in credits.
ParallelParallel Search becomes available in AWS Marketplace
Parallel said its Search API was available in AWS Marketplace for use with Amazon Bedrock models and as an MCP server in Bedrock AgentCore. The company said teams using AWS can apply eligible purchases toward their AWS spend commitment.
Parallel# Introducing the Scale plan for Parallel
Scale gives teams 1,500 requests/min on Search and Extract and ZDR for $1,000 a month, all of it spent as API credits.
parallel.ai announcement