What happened
Parallel said GPT-6 Astra researched six labor-market statistics across four states over six months in half the time of prior models, with roughly 50% code cost reduction and the same research quality.
Why it matters
Doing the same research in half the time at roughly half the code cost means Parallel can divide work among multiple agents instead of running one long search sequence.
What to watch
The gains are measured on Parallel's own longest-running research tasks, so whether they hold for other workloads is undecided. Watch the 50% code cost reduction as the figure Parallel cites.
WHO IT HITSDevelopers building AI research agents — particularly teams doing long, multi-source research for financial and legal customers — could get the same quality answer with fewer research calls and less token spend, based on Parallel's test.
Summaries like this, in your inbox every morning.
Parallel's business is developer infrastructure for AI agents that do knowledge work over the web, spanning web grounding for voice agents and research for financial institutions and legal customers. Its longest-running research tasks had typically required a bigger model with extended reasoning, which consumed more time and resources. The test with GPT-6 Astra focused on exactly that kind of task: an agent compiling six months of labor-market data across four states into a single report.
Beyond the headline time and cost figures, Parallel observed a change in how the agent searched. According to Devin Gupta, a Member of Technical Staff at Parallel Web Systems, Astra issued more targeted search queries and stayed focused on the ultimate task, incorporating its world knowledge compared with previous models. The company also noted that Astra can delegate specific research tasks to sub-agents, letting work happen simultaneously rather than moving through a single sequence of searches.
Taken together, these results point to a more practical way for Parallel to split demanding research across multiple agents without losing quality — though the reported improvements come from Parallel's own testing on its longest-running tasks, so how far they extend to other workloads is likely to depend on further results.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, LINE, or Slack.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Identity Digital spun out Known Systems AI Inc., based on its Innovation Labs unit and DNSid, a 'birth certifi…
Qualcomm and MediaTek announced on the same day that they are entering the Googlebook supply chain, as preorde…
Brahma AI raised $150 million through preferred shares, with $100 million from Multiples Alternate Asset Manag…
Singapore's electronics exports rose 132% from a year ago in August, with personal computer shipments up 238%…
A Bloomberg gauge of 30 Chinese technology stocks with the biggest overseas revenue exposure returned 36% this…
China's National Internet Information Office has opened an investigation into Moonshot AI and DeepSeek over su…
