ChatGPT Can Now Split One Job Across Several Agents at Once: The First Research Task to Hand It This Week
On October 16, OpenAI gave GPT-6 Pro in ChatGPT parallel agents: one task, several agents, each working a different part at the same time. Here is the one research job worth testing it on, the brief that makes the split work, and how to run the same play without Pro.
What happened: On October 16, 2026, OpenAI upgraded GPT-6 Pro in ChatGPT. Responses are faster, it uses plugins and tools better, and it can now split one task across several agents that each work a different part at the same time. You watch each agent's progress in the conversation and open each finished piece to see what it did. It is rolling out to Pro users in the Chat tab. In company-managed workspaces, availability depends on the rollout and your admin settings.
What OpenAI has not said yet: How many agents one task can use, how parallel work counts against your Pro allowance, or any quality numbers. There is no API version so far. Treat the first week as a test.
The so-what: The best fit is a research job that normally sprawls across a dozen browser tabs. Five competitors' prices and reviews. Four software tools you are choosing between. Three suppliers' terms. Each part stands on its own, so splitting it saves time, and each piece comes back separately so you can check it.
The move: give it one research job this week. Pick a decision you are making in the next 30 days that needs outside facts. Write the parts out yourself, because a clean split is what makes the merged answer trustworthy. Paste this:
ROLE: You are a research team lead. Run each part below as its own agent, in parallel.
THE DECISION THIS SUPPORTS: [one sentence, e.g. which scheduling and invoicing software we switch to by December]
PARTS (one agent each):
1. [e.g. Tool A]
2. [e.g. Tool B]
3. [e.g. Tool C]
4. [e.g. Tool D]
EACH AGENT RETURNS THE SAME COLUMNS: [e.g. published price for 10 users, contract length, the 3 most common complaints in reviews from the last 12 months, whether it connects to QuickBooks]
Plus a Sources line with a link for every number, and a Could Not Find line.
THEN: Merge everything into one comparison table. Below it, list the 3 facts that should matter most to my decision and why. Mark any number without a source as UNVERIFIED.
CONSTRAINTS: Public information only. Do not log in to anything, fill out forms, start trials, or contact anyone. If a part cannot be finished, say so instead of guessing.
Then check the work: Open each completed agent task. Click one source link per agent and confirm the number on the page matches the table. If one agent's sources are thin, rerun only that part.
Not on Pro? Run the same brief by hand. One chat per part, then paste all the results into a fresh chat with the THEN section. Slower, same structure, and it works in Claude or any ChatGPT plan.
Keep it safe: Start with public information only. Agents that use plugins and tools act on whatever you have connected, so keep your inbox, client files, and payment tools disconnected for this test. A research brief never needs client data in it.
What it's worth: A competitor or vendor sweep that eats an afternoon of tabs can become a 15-minute review of finished tables you can actually verify.