How to Use ChatGPT GPT-6 Pro Parallel Agents in the Chat Tab

Introduction

On 16 October 2026, OpenAI turned the ChatGPT Chat tab into a multi-agent workspace for Pro subscribers. GPT-6 Pro can now split one request across several agents that work on different parts of the task at the same time, while you follow each agent’s progress and open each completed task to review what it did. This tutorial shows you how to check whether you have access, how to write a task brief that splits cleanly, and how to merge the agents’ work without letting a quiet mistake slip through. It also marks everything OpenAI has not told us yet, because those gaps decide whether parallel agents save you time or spend your weekly allowance.

Quick answer

Pick GPT-6 at Pro effort in the Chat tab, send one task brief that names independent parts with a deliverable each, watch each agent’s progress panel, open every completed task, and reconcile contradictions before merging.

Evidence

The paper trail for this feature is short, and that is part of the story. OpenAI announced the change on 16 October 2026 in a post on X and in the ChatGPT release notes, and the release note is four lines long. Under the heading “Updates to GPT-6 Pro model in ChatGPT”, the ChatGPT release notes state that GPT-6 Pro in the chat tab now offers faster responses, improved use of plugins and tools, and “multiple agents that can work on different parts of a task in parallel. You can follow their progress in the conversation and open completed tasks to see what each agent did.” That sentence is the entire product description. OpenAI publishes no latency figures, no agent count, and no statement on how a parallel task counts against usage limits.

The naming is already inconsistent, so pin it down before you search for help. OpenAI’s first X post said “GPT-6 Pro” and was edited about half an hour later to “GPT-6 Astra Pro”. The release notes and the Help Center keep calling it GPT-6 Pro. They are the same thing: the Pro reasoning option in Chat, powered by GPT-6 Astra, the frontier model OpenAI released on 3 September 2026.

Access is the first hard limit. OpenAI’s Business models and limits guide lists GPT-6 Pro, powered by GPT-6 Astra, for Pro $100, Pro $200, Business, and Enterprise plans, and the October 16 note says the update is rolling out to eligible Pro users, with managed workspaces depending on their settings. Plus subscribers do not get GPT-6 Pro in Chat at all, although they do get GPT-6 Astra inside ChatGPT Work and Codex.

The second hard limit is the allowance, and it is tighter than the marketing suggests. GPT-6 Pro is limited on every plan that has it. On Pro $200 the allowance is 200 messages a week, and on Pro $100 it is 50 a week, in both cases shared with GPT-5.6 Sol Pro. Business Standard includes 15 Pro messages a month, and Business Premium includes 50 a week, again shared across the two Pro models. For scale, OpenAI’s enterprise token rate card prices GPT-6 Pro at $10.00 per 1M input tokens and $50.00 per 1M output tokens, the same row as GPT-6 Astra, which confirms that Pro effort is the most expensive way to run a ChatGPT conversation.

The only measured speed data comes from independent testing, not from OpenAI. One tester ran comparable tasks sequentially and with three parallel agents: the sequential run took 4 minutes 51 seconds, the parallel runs averaged 4 minutes 25 seconds, about 9 percent faster, but one parallel run came in 10 seconds slower than the control and another 62 seconds faster. Treat parallel agents as a way to split work, not as a guaranteed speedup. The gain is real but small and inconsistent, and it only appears when the task genuinely splits.

Prerequisites

• A ChatGPT Pro ($100 or $200 a month), Business, or Enterprise account. Free, Go, and Plus accounts do not include GPT-6 Pro in the Chat tab.
• Access to the Chat tab on web or desktop, with GPT-6 visible in the model menu. The rollout is gradual, so the feature may not appear on your account yet.
• A task that splits into two or more parts that do not depend on each other’s output. Research briefs, multi-section drafts, and comparison tables split well. A single linear argument and a debugging session do not.
• A note of your remaining weekly allowance before you start, because OpenAI has not documented whether one fanned-out task counts as one message or several.

Step-by-step guide

Step 1. Confirm GPT-6 Pro is available on your account

Open ChatGPT, start a new chat in the Chat tab, and open the model menu. Select GPT-6, then set the reasoning effort to Pro. If Pro is missing, you do not have access yet. Confirm your plan in Settings first, because Plus accounts never see GPT-6 Pro in Chat. On a Business or Enterprise workspace, ask your admin whether the GPT-6 Astra model permission is enabled, since managed workspaces depend on their settings and rollout eligibility.

Step 2. Pick a task that actually splits

Parallel agents help when the parts are independent and each part has a checkable output. Good candidates are a research brief with three separate sub-questions, a comparison of four tools where each tool gets its own section, and a document where the introduction, the data section, and the FAQ can be drafted separately. Bad candidates are a task where every step depends on the previous one. If you cannot name the parts and their deliverables in one sentence each, the task does not split yet.

Step 3. Write a brief the agents can split on

OpenAI has not documented how GPT-6 Pro decides to split a task, so the brief does the work. Write the goal first, then the parts, then one deliverable per part, then the rules that keep them from colliding. Use a template like this:

Goal: produce a 1,500-word comparison of three note-taking apps
for a team that works offline half the week.

Part 1: App A. Deliverable: 400 words covering offline sync,
pricing, and export formats. Cite the vendor docs.
Part 2: App B. Same deliverable, same length, same sources rule.

Rules: use only the vendor's current pricing page for prices.
Do not invent features. Flag anything you could not verify
instead of guessing.

The rules matter more than the parts. A shared source of truth for prices, a ban on guessing, and a requirement to flag unverifiable claims stop the most common failure mode: two agents confidently reporting different prices for the same product. Tell the agents you will assemble the final document yourself, so they do not each write their own conclusion.

Step 4. Send the brief with GPT-6 Pro selected

Send the brief in the Chat tab with GPT-6 at Pro effort selected. The response streams in the conversation, and each agent’s progress appears as its own activity. You do not pick the number of agents, and you do not assign parts to agents yourself. GPT-6 Pro decides the split, so you cannot force a three-way split on a task the model judges to be a single piece of work. If the response comes back as one sequential answer with no agent activity, the model did not split your task. Rewrite the brief with sharper part boundaries and try again, or accept that this task runs sequentially.

Step 5. Follow progress and open each completed task

Watch the progress indicators while the agents run, and open each completed task when it finishes instead of waiting for a single final answer. Each completed task is your review checkpoint, and it is the only place where you can see what one agent actually did before its output is blended into the whole. Read each part against its deliverable: did it answer the sub-question, did it cite a source you can open, and did it flag anything it could not verify? If one agent’s part is thin, ask a follow-up in the same conversation that narrows that part only. Do not re-send the whole brief, which starts a new split and spends more of your allowance.

Step 6. Merge the parts and verify before you use them

Assemble the final document yourself, and verify it more carefully than a single-agent answer. Check that prices and dates match across parts, that no two parts contradict each other, and that nothing is duplicated because two agents covered the same ground. Then ask GPT-6 Pro to list every factual claim in the merged document that lacks a cited source. Parallel work multiplies output and the surface where an error can hide.

Step 7. Track what the task cost you

After the task, check your remaining GPT-6 Pro allowance in Settings. OpenAI has not documented whether a task fanned out to three agents consumes one message or three, so watch the counter before and after a real task to learn the cost. If it drops by more than one, budget future parallel tasks accordingly, and keep your longest ones for the start of the week when the allowance resets.

Troubleshooting

The Pro option is missing from the model menu

You are either on a plan that does not include GPT-6 Pro, or your account is not in the rollout yet. Confirm the plan in Settings first: Free, Go, and Plus never see GPT-6 Pro in Chat. If you are on Pro, Business, or Enterprise and still do not see it, wait. The rollout is gradual and account-based, and no setting forces it. Business and Enterprise users should also check with their workspace admin, because the GPT-6 Astra permission can be off by default in some workspaces.

The model answers sequentially instead of splitting the task

GPT-6 Pro decides whether to split, and it splits only when the brief makes the parts obvious. Rewrite the brief so each part has its own deliverable and its own sources, and remove any sentence that ties the parts together. If the task genuinely depends on sequential reasoning, accept the sequential answer. Forcing a split on a dependent task produces worse output, not faster output.

Two agents report different facts for the same thing

This is the most common real failure, and it comes from a brief without a shared source of truth. Stop and reconcile before you merge anything. Open both completed tasks, compare the sources each agent cited, and ask GPT-6 Pro to decide which source is current, with a link. Then add a rule to your next brief naming the single authoritative source for the disputed fact. Do not pick the answer that sounds more confident.

One agent’s part comes back thin or empty

A part with a vague deliverable gets a vague result. Reply in the same conversation with a narrower instruction for that part only, naming the exact output you want and the source to use. If the part stays thin after a second attempt, draft that section yourself from the sources the other agents cited. A run with two strong parts and one weak part is still a net win if you spot the weak part, which is why Step 5 says to open each completed task.

You hit the weekly GPT-6 Pro limit mid-task

The allowance is shared with GPT-5.6 Sol Pro, so heavy Pro use earlier in the week leaves less room for a parallel task. When the limit is reached, ChatGPT switches to a lower effort model. Independent reporting describes the fallback as GPT-5.6 Thinking at Medium on Pro $200, but OpenAI’s own docs do not name it, so confirm what your account actually switches to. The practical fix is planning: run parallel tasks early in the week, keep the brief tight so a retry costs less, and save a copy of the partial output before the limit hits, because a task interrupted by the allowance does not resume where it stopped.

FAQ

Which ChatGPT plans get GPT-6 Pro parallel agents?

Parallel agents live inside GPT-6 Pro in the Chat tab, and OpenAI’s models guide lists GPT-6 Pro for Pro $100, Pro $200, Business, and Enterprise plans. Plus does not include GPT-6 Pro in Chat, although Plus includes GPT-6 Astra inside ChatGPT Work and Codex.

How many agents can work on one task at once?

OpenAI has not published a number. The release note says only that multiple agents can work on different parts of a task in parallel.

Do parallel agents make answers faster?

Sometimes, and not by much. One independent test found parallel runs averaged 4 minutes 25 seconds against 4 minutes 51 seconds sequential, about 9 percent faster, but one parallel run was 10 seconds slower than the control. OpenAI publishes no speed figures for this feature.

Does one parallel task use one GPT-6 Pro message or several?

OpenAI has not said. GPT-6 Pro carries a usage allowance, 200 messages a week on Pro $200 and 50 a week on Pro $100, shared with GPT-5.6 Sol Pro. Until OpenAI documents how a fanned-out task counts, assume it costs more than one.

Can I use parallel agents in ChatGPT Work or Codex instead of Chat?

No, not this feature. The October 16 announcement covers GPT-6 Pro in the Chat tab only. ChatGPT Work and Codex have their own parallel mechanisms, but that is a different surface.

What should I do when two agents contradict each other?

Treat the contradiction as the signal, not as noise. Open each completed task, compare the evidence each agent cited, and ask GPT-6 Pro to reconcile the two with sources before you accept either version.

Conclusion

GPT-6 Pro parallel agents are a small, honest upgrade, not a new way of working. The feature splits a well-briefed task across several agents, shows you each agent’s progress, and lets you open each completed task before anything is merged, and that visibility is the real improvement. The limits matter as much as the capability: Pro plans only, a weekly allowance that OpenAI has not explained how parallel tasks consume, no published agent count, and a speed gain that independent testing measured at about 9 percent. Use parallel agents for research and drafting work that splits into independent, checkable parts, brief them with a shared source of truth, review every completed task, and merge the result yourself. For a comparison point, Anthropic shipped a similar idea on 10 October with Claude Code Projects, where one conversation directs parallel cloud threads with shared memory, and our Claude Code subagent effort levels guide shows how to run low-effort scouts next to high-effort reviewers. For more ChatGPT automation, see how to use MCP events in ChatGPT and the complete n8n guide.

Frequently asked questions