Parallel cut research time and cost in half with GPT‑6 Astra

Parallel, a company building developer infrastructure for AI agents performing knowledge work over the web, has reported significant performance improvements using OpenAI's GPT-6 Astra. Previously, handling Parallel's longest-running research tasks required using larger models with extended reasoning, which consumed more time and resources. In a recent test evaluated by Parallel Member of Technical Staff Devin Gupta, an agent powered by GPT-6 Astra was tasked with researching six different labor-market statistics across four states over a six-month period, involving multiple websites and a comprehensive final report. According to the company, GPT-6 Astra completed the work in half the time of prior models while reducing code costs by approximately 50 percent, all without sacrificing research quality. Parallel observed that GPT-6 Astra executed more focused searches and reached useful results in fewer steps. Furthermore, the model's increased efficiency has made it more practical to divide workloads among multiple sub-agents simultaneously, bypassing the limitations of a single sequential search process. Parallel stated that these enhancements provide a better path from complex questions to thoroughly researched answers, lowering costs and enabling the handling of more demanding research tasks at scale.