GPT-6 Astra cuts Parallel's research time and cost in half
AI developer infrastructure company Parallel reports that GPT-6 Astra halved both the time and cost of complex, multi-source labor-market research tasks compared with previous frontier models, while maintaining output quality and enabling more efficient multi-agent workflows.
Parallel, a company building developer infrastructure for AI agents that perform knowledge work over the web, has reported cutting research time and cost by roughly half after switching to GPT-6 Astra, according to tests run by the company and published on 24 September 2026.
In a benchmark designed to stress its systems, Parallel tasked an agent with researching six different labor-market statistics across four states over a six-month period. The agent was required to search multiple websites, gather the relevant information, and synthesize it into a single research report. According to the company, GPT-6 Astra completed the work in half the time of prior models, with a roughly 50% reduction in code cost, while delivering the same quality of results.
Devin Gupta, Member of Technical Staff at Parallel Web Systems, attributed the gains to the model's efficiency. "With Astra, we've demonstrated that you can get the same high-quality research much, much faster with fewer research calls and less tokens," Gupta said.
Beyond speed, Parallel found that GPT-6 Astra pursued more focused search paths. Rather than issuing broad, sequential queries, the model targeted its searches and drew on its built-in world knowledge to stay on task. "Astra issued more targeted search queries and focused on the ultimate task better, incorporating its world knowledge compared to previous models," Gupta said.
That change in search behavior also affected how Parallel structures its multi-agent workflows. The company found it more practical to divide research tasks among sub-agents, with GPT-6 Astra delegating specific work so that tasks could proceed simultaneously rather than in a single long sequence. According to the source report, this shift from sequential to parallel execution removes a significant bottleneck for teams building AI agents that carry out heavy web research.
No independent verification of Parallel's test results was available at the time of publication, and the findings represent the company's own internal benchmarks.
Prepared with AI assistance and reviewed by the editorial team.
