Parag Agrawal’s New AI Firm ‘Parallel’ Outperforms GPT-5 and Humans in Web Research

Parag Agrawal’s New AI Firm ‘Parallel’ Outperforms GPT-5 and Humans in Web Research

Parag Agrawal- the former Twitter CEO, who was fired by Musk during his chaotic takeover in 2022, is back in the game. This time, with an ambitious AI venture aimed at reshaping how machines use the internet. 

Agrawal has co-founded Parallel, a Silicon Valley startup that develops infrastructure that allows AI systems to navigate and extract information from the web better than humans or existing AI models in the market. Parallel claims its technology outperforms not only OpenAI’s GPT-5 but even trained human researchers on complex, multi-step research tasks.

From Twitter’s Power Struggles to AI Infrastructure

Agrawal, who served as Twitter’s chief technology officer before taking over as CEO in late 2021, faced a high-profile and abrupt departure less than a year into the role. His exit, along with several other executives, came just hours after Elon Musk confirmed his controversial $44 billion acquisition of the microblogging platform.

After maintaining a relatively low profile for nearly three years, Agrawal returns to tech’s centre stage with a startup addressing what he and his team call “the web’s second user”: artificial intelligence.

“Humans browse the web. Machines don’t,” the company’s messaging reads. “We're building the infrastructure for AIs to use the web the way humans never could.”

Parallel’s flagship offering is an API that allows AI agents to conduct deep web research tasks in real time. Unlike conventional search engines built for humans, Parallel’s tools are optimised for machine-to-machine interaction, enabling agents to read, retrieve, and reason across thousands of documents in seconds.

On BrowseComp, a test designed by OpenAI to measure multi-hop reasoning across live web data, Parallel scored 58% accuracy, outperforming the newly launched GPT-5 and even trained human researchers who were given two hours per task (25%).

On another benchmark, DeepResearch Bench, which involves expert-level queries across domains like finance, science, and software development, Parallel again topped GPT-5  achieving an 82% win rate compared to GPT-5’s 66%.

Unlike traditional search platforms, Parallel delivers structured, machine-ready outputs. It can be used to automate everything from coding assistance to claims verification and investment analysis.

Parallel is backed by Silicon Valley names, including Khosla Ventures, First Round Capital, Index Ventures, and others. Its founding team includes veterans from Google, Twitter, Stripe, Airbnb, and Waymo, many of whom worked on large-scale infrastructure and applied machine learning systems.

Though still early in its journey, Parallel already claims to be powering millions of research tasks daily.

Read More: How Intel Insider Varun Gupta Stole Trade Secrets For Microsoft