GPT-6 Pro's Parallel Agents: What ChatGPT's Multi-Agent Mode Does

GPT-6 Pro's new parallel agent mode lets ChatGPT split one task into pieces, hand each piece to a separate agent, and work on all of them at the same time instead of grinding through one step after another. OpenAI shipped it on October 16, 2026, as an upgrade to GPT-6 Astra Pro (most people still call it "GPT-6 Pro" out of habit) inside ChatGPT's Chat tab. Right now it's locked to Pro subscribers, and OpenAI hasn't said much about how it actually works under the hood.

I've been testing it for a couple of days and reading everything OpenAI, and the people complaining about OpenAI, have said about it. Here's what's real, what's missing, and how it stacks up against the subagent setup Claude Code users already take for granted.

What does GPT-6 Pro's parallel agent mode actually do?

The official line, straight from OpenAI's announcement, is that "multiple agents can tackle different parts of a task in parallel. You can follow each agent's progress and open completed tasks to review their work."

In practice that means:

• Ask ChatGPT something big and multi-part (plan a trip, research five competitors, draft three versions of a pitch) and it can break the request into subtasks
• Each subtask runs as its own agent, visible in the chat as a separate progress thread
• You can click into any agent that's finished to see what it produced before the whole thing wraps up
• The same update brought faster responses and better use of plugins and tools generally, separate from the parallel piece

That's it. That's the whole public description. No agent count limit, no explanation of how OpenAI decides what gets split versus handled by one agent, and no mention of what happens when two agents produce conflicting answers.

Who actually gets access to it?

Parallel agents are rolling out to ChatGPT Pro accounts in the Chat tab. OpenAI quietly restructured Pro into three tiers this month: Pro 100 ($100/month), Pro 200 ($200/month, the plan most existing Pro users are on), and a new Pro 500 ($500/month) that bundles faster model variants. The parallel agent feature doesn't appear to require the top tier, just a Pro subscription of some kind.

Business and Enterprise access depends on workspace settings, and OpenAI hasn't given a date for when (or if) it reaches Plus users. If you're on the $20/month Plus plan, you're watching this one from the sidelines for now.

Is this actually new, or did OpenAI just expose something that already existed?

This is the part most coverage glossed over. According to OpenAI's own developer docs for GPT-6 Astra, the model already "supports multi-agent orchestration and can delegate work to subagents for parallel execution" at the API level, it's just that developers had to prompt it to do so. The ChatGPT feature is that same capability wrapped in a chat UI with a progress tracker, turned on by default instead of requiring a prompt.

So no, nobody invented parallel agents on October 16. OpenAI took a capability that was already sitting in the Astra model and made it visible to regular Pro subscribers without them needing to write a system prompt for it.

What's missing from OpenAI's announcement?

Quite a lot, honestly. Coverage from RuntimeWire and Mindverse both flagged the same gap: there are no latency figures, no benchmarks, and no examples showing the speed or quality gain over the old one-step-at-a-time approach. Mindverse put it bluntly, calling out unanswered questions around agent coordination, tool permissions, and administrative availability.

That's a real problem if you're trying to decide whether this is worth anything for your workflow. "Trust us, it's faster" isn't a spec. Until OpenAI publishes real numbers, treat every claim about speedup as marketing until you've run your own side-by-side test.

How does this compare to Claude Code subagents?

This is where the comparison gets interesting, because Anthropic's version of this idea has existed for a while and works completely differently.

The gap that matters most to me is configuration. With Claude Code, you write a subagent that only has read access for a research pass, and a different one with full edit permissions for implementation, and you know exactly what each one can touch. With ChatGPT's parallel agents, you have no visibility into what each agent is allowed to do or how it decided to split the work in the first place. You just watch the progress bars.

I get why OpenAI built it this way. Most ChatGPT users don't want to write YAML frontmatter for a subagent, they want it to just work. But "just works" also means "you can't inspect it," and for anything beyond casual research tasks, that trade bothers me. If you've used Claude Code's nested sessions to run real parallel work on a codebase, going back to a black-box version feels like a downgrade in control even if it's faster on paper.

Should you pay for ChatGPT Pro just for this?

No, not yet. If you're already paying $200/month for Pro, turn it on and see if it saves you time on the kind of sprawling, multi-part requests you already send ChatGPT. But I wouldn't upgrade from Plus, and I definitely wouldn't treat this as a reason to switch away from a coding-specific tool if your work is actually software engineering. For that, the subagent model in Claude Code is still the more transparent, more controllable option, and it doesn't cost you a $200/month tier to use it.

The honest read is that this is OpenAI catching up to a pattern that agentic coding tools normalized months ago, repackaged for a general audience that's never heard of a subagent. That's a fine thing to ship. It's just not the leap the announcement tweet makes it sound like.

If you're curious what GPT-6 Astra can actually build when you point it at a real project instead of a chat request, I wrote about that in building a game with GPT-6 Astra. And if you want the fuller picture of how OpenAI's agent features in ChatGPT work outside of this update, ChatGPT's work-focused Agent mode covers the other half of OpenAI's agent push. For the Claude side of this comparison, Claude Code's nested desktop sessions is the closest thing Anthropic has shipped to a visual multi-agent view.

Watch the announcement

OpenAI's post is short, two sentences and a screenshot, which tells you most of what you need to know about how much detail they're sharing at launch.