Parallel Web Systems

September 16th, we're hosting a happy hour in San Francisco after Dreamforce. Join us @ Shoji for Japanese spirits and a gathering of the people building today's most forward-looking AI products. 5:00-7:00 PM. RSVP here: https://lnkd.in/eSic2fMG

The bad news: you’re probably evaluating web search wrong. The good news: our team wrote a guide to help you and your agents do it right. Our main takeaway: don’t ask "Which search provider is best?" Instead, ask: “Which model × search configuration is best for this task, at the cost and latency we can afford?”

GLM-5.3-Flash is here. We benchmarked it against GPT-5.6 Luna with a suite of different web search solutions to see just how different the cost and performance would be. The results were stark: - GLM-5.3-Flash w/ Parallel Fast performed better on cost per task (~7x) and accuracy vs. GPT-5.6 with built-in web_search (11 pts). - GPT-5.6 Luna w/ Parallel Fast delivers equal accuracy (76%) but at twice the cost vs. GLM 5.3 ($0.02 per task).

Benchmark Highlight: Parallel Web Systems Benchmark: Company Funding Enrichment Openbenchmarks put 20 endpoints from 16 providers: Input: Company-Domain Output: Latest Funding Round Judged against source-verified rounds. We measured Parallel's two endpoints - the Task API and the Responses API Here's what we found Best accuracy per dollar at the top of the board (on self-serve version): Parallel's Task API got the latest stage right 90.0% of the time at $0.025 per run for backfill enrichment task. At 90%+ accuracy, nothing else comes close on cost or speed. Cheapest agent API on both boards. $25 per 1,000 Task runs. The other two long-running agent endpoints measured bill 4× and 5.6× that per call, for the same question and the same output schema. It beat the incumbent datasets: 90.0% on established rounds, against the two legacy funding datasets on the board at 85.8% and 71.2%. 100% company identification and resolution: Not one traditional GTM data provider managed it: the best of that group resolved 97.7%, and three of them failed to identify a quarter to half the list. This is important for higher CRM enrichment rate. Parallel is the only provider on the benchmark with two separate endpoints at 89.0% or better on the enrichment board. You pick the shape that fits your latency and cost budget instead of taking whatever the one endpoint gives you.

Watch Parag Agrawal break down our company thesis on the latest episode of Training Data with Sonya Huang and Andrew Reed: https://lnkd.in/eD4y_k7h

Parallel's full API suite is now available in the Google Cloud Marketplace! Businesses can now quickly and easily provision Parallel's best-in-class web search tools and agents directly through Google Cloud, with consolidated billing and committed spend drawdown. Get started: https://lnkd.in/eeqDujMQ

Your web search shouldn’t cost more than your inference. It's simple: If you’re using GPT 5.6 Luna or similarly priced models, chances are you’re overpaying for search. Save on web grounding with the Parallel Search API: https://lnkd.in/eb72UPAw

A little about the significance of a $1 CPM, 5-10x more affordable web search product for AIs and agents. 45 days ago we launched a Turbo search mode. Our goal was: how do we get people to think about web search as a no-brainer always-on part of AI and agentic workflows? For many use cases, adding web access can 2-3x eval results! Turbo was unique: very fast (200-300ms) and the first ever $1 CPM AI-focused search SKU. Comparable offerings (including our own) are normally $5-$10 CPM. We felt both would encourage use, but we didn't know which would matter more. Since then, we kept hearing one piece of feedback above all others: "frontier intelligence cost is dropping, and $1 CPM for search makes it a no-brainer for my use case, but I wish I could trade some of Turbo's speed for even higher quality." If frontier intelligence is 5x cheaper than it was a year ago, why isn't search? But there was a reason why deeper search was priced $5+. The search index is large, and growing. Users want fresh data and long tail data. Getting accurate, relevant, and succinct summaries for LLMs to consume requires inference over large parts of the web, often at request time. This costs time and money (compute, inference, index, ...). $1 CPM for frontier quality web grounding felt hard to achieve. But the directive was clear. Parallel Fast mode is the culmination of that effort. It is $1 CPM, just like Turbo, meaning it's a tiny fraction of the total cost of inference, even on cheap models. It's consistently sub-1s (won't add meaningful latency to any LLMs). And most importantly, it's frontier intelligence. No tradeoffs. It's ranked #3 on Artificial Analysis, only 2 points (73 vs 75) below the best search product in the world (our own Advanced mode). Trivially outperforming models at 5-10x the price. For giving AIs and agents web access, we think we created the most obvious default search product. And while we're excited about getting existing search users over to Fast, more importantly, we think it will make instant web access a no-brainer for many more developers.