Parallel builds developer infrastructure for AI agents that perform knowledge work on the web, supporting tasks from web grounding for voice agents to research for financial and legal customers. Its tools combine frontier models with web search to produce answers.
For Parallel’s longest‑running and most complex research tasks, achieving high‑quality answers traditionally required using a larger model with extended reasoning, which increased time and resource consumption. Parallel reports that GPT‑6 Astra has produced a marked improvement in these metrics.
Test task and results
In an internal test, Parallel asked its agent to research six different labor‑market statistics across four U.S. states over a six‑month period. The task required searching multiple websites, collecting data, and compiling the findings into a single research report.
According to Parallel, GPT‑6 Astra completed the work in roughly half the time of prior models while reducing code costs by about 50%, and delivered research of the same quality as earlier approaches.
More focused searches and fewer steps
Parallel observed that GPT‑6 Astra executed more focused searches and required fewer steps to reach a useful result. In practice, the model more efficiently identified relevant sources and synthesized information more quickly.
Parallelization and delegation to sub‑agents
The improved efficiency also made it practical for Parallel to split research across multiple agents. GPT‑6 Astra was able to delegate specific research tasks to sub‑agents, enabling concurrent work and shortening the time spent progressing through a single sequence of searches.
Impact on workflows
Parallel says the model provides a clearer path from a complex question to a researched answer: less waiting, lower costs, and greater capacity to handle demanding research at scale. These improvements are particularly relevant for organizations that regularly perform multi‑source, time‑intensive data collection and synthesis.
(The figures and observations reported here come from Parallel’s internal tests; the company did not disclose precise operational timelines or additional implementation details.)
