In a single customer test, OpenAI says GPT-6 Astra halved Parallel’s web-research time and cut code-related costs by about 50%.
OpenAI says GPT-6 Astra completed a long-running web-research task for startup Parallel in half the time required by earlier models, while reducing code-related costs by roughly 50%. Parallel judged the resulting research to be equivalent in quality to its previous approach. The figures come from one specified customer workflow, not a broad benchmark.
Parallel builds web-based infrastructure for AI agents by combining advanced models with web search. OpenAI identifies use cases including supplying current context to voice agents and conducting research for financial and legal customers. According to the company, Parallel had previously depended on larger models with extended reasoning, which consumed more time and computational resources.
The test asked a Parallel agent to investigate six labor-market statistics across four US states over six months. It had to search multiple websites, extract the relevant data and consolidate its findings into a single report. OpenAI says Astra finished in half the earlier time while maintaining quality as assessed by Parallel.
OpenAI attributes the efficiency gain to more targeted searches and fewer steps before reaching a useful result. Parallel said Astra remained better focused on the final objective and used its existing knowledge more effectively. Neither absolute completion times nor costs in dollars were disclosed, so the reported improvement cannot be converted into a concrete time or budget saving from the published information.
OpenAI also says the lower workload makes it more practical for Parallel to divide research among multiple sub-agents. Separate assignments can then run concurrently instead of moving through one sequential search chain, potentially expanding how often Parallel uses parallel execution within its system.
The evidence remains narrow: this is a customer case study covering one labor-market research task. OpenAI does not name the comparison models, define the quality-assessment method, disclose search or token counts, or provide statistical testing. It also makes no claim that the approximately 50% improvements apply to other customers, industries or configurations. The case therefore supports a workflow-specific result, not a universal performance advantage for Astra.
Sources
Originally published by AI Feed: OpenAI Says GPT-6 Astra Halved Parallel's Research Time.
Top comments (0)