ChatGPT GPT-6 Astra Pro makes multiple agents work in parallel on a single task

GPT-6 Astra Pro: ChatGPT now runs multiple agents in parallel on a single task

🔎 Multi-agent leaves the terminal and enters your chat

On October 16, 2026, OpenAI quietly moved ChatGPT into the multi-agent era. An update to GPT-6 Astra Pro — called "GPT-6 Pro" in the official release notes, "GPT-6 Astra Pro" on X — now allows a single request to be split into multiple agents working simultaneously, each one visible and individually inspectable. The OpenAI announcement sums it up in its own title: parallelism reduces time and cost.

Until now, getting multiple AIs to collaborate on a single task was a developer's DIY affair: orchestration frameworks, homemade scripts, projects like Orca. OpenAI has just put this mechanism into the mainstream product, without a single line of code to write.

Should we hail it as a revolution? The first independent test, published the same day by Kingy.ai, calls for caution: roughly a 9% speed gain, no measured accuracy advantage, but a real change in how you inspect an AI's work. Here's what we know, what we've measured, and what OpenAI isn't saying yet.

The Essentials

• The update: on October 16, 2026, OpenAI enables parallel agents in ChatGPT for eligible Pro users, with "faster responses" and "better use of plugins and tools".
• The principle: a request splits into multiple agents that work simultaneously; progress is displayed in the conversation and each completed task opens so you can review its work.
• The numbers: in Kingy.ai's test, 3 agents in parallel complete a brief in 4 min 25 s on average versus 4 min 51 s sequentially — 26 seconds saved, or about 9%.
• The real gain: inspection. Three named agents, three viewable panels with sources and caveats.
• The blind spots: no documented maximum number of agents, no per-run quota counter, no quantified benchmark to back up the "faster" claim.
• Access: nothing to enable, just select GPT-6 Pro level in the chat (Pro 100, Pro 200, Business and Enterprise plans).

Recommended tools

What This Update Really Changes

Direct answer: ChatGPT no longer just answers — it delegates. A request can now be split across multiple agents working at the same time, and you can open each one's work once the task is complete.

ChatGPT's release notes sum up the update in one sentence: "Faster responses, improved use of plugins and tools, and multiple agents that tackle different parts of a task in parallel". On OpenAI's X post, the wording is identical, with one clarification: the rollout applies to eligible Pro accounts, in the Chat tab.

Concretely, you track each agent's progress directly in the conversation, and every completed task opens up so you can review its work. EconoTimes confirms a gradual rollout starting October 16 for eligible Pro users, with progress display and results viewable agent by agent.

One small detail that reveals the surrounding confusion: even the naming can't make up its mind. The release notes say "GPT-6 Pro", the X post says "GPT-6 Astra Pro". Same update, two names — a sign that the communication ran faster than the documentation.

A Rollout With No Switch

On the access side, it's disarmingly simple: nothing to activate. According to vibecoding, the feature ships with the Pro 100, Pro 200, Business, and Enterprise plans — you just need to select GPT-6 at the Pro level in the chat. The limit consumed by a parallel run, however, is not documented.

A caveat, though, for managed workspaces: availability there depends on workspace settings and rollout eligibility. If you work in a company, the decision belongs to your administrator, not to you. The OpenAI community confirms this eligibility-based mechanism.

Useful context: this update follows in the footsteps of Astra's "dots" agents, which stayed active in the background — and which had been excluded from the EEA, Switzerland, and the UK at launch, according to GoldenEra. The geographic coverage of the parallel agents is not detailed. As a reminder, the model itself scored 64.6% on Terminal-Bench, with standout results in science and on ARC-AGI-3 (our analysis).

The numbers from the first independent test: +9%, no miracle

Direct answer: the speed gain is real but modest — 26 seconds out of nearly five minutes, or about 9%.

Kingy.ai published the very same day the first independent test. Protocol: two sequential/parallel pairs on a frozen brief, split into 3 lanes — three distinct work tracks. Average sequential result: 4 min 51 s. In parallel with 3 agents: 4 min 25 s. Average gain: 26 seconds, about 9%.

The per-run detail is more instructive than the average. The first parallel run was 10 seconds slower than its sequential control (4 min 56 s vs 4 min 46 s). The second was 62 seconds faster (3 min 54 s vs 4 min 46 s). In other words: one run in the red, one run deep in the green, a lukewarm average.

Variance eats the promise

A 72-second spread between two runs of the same setup, for an average gain of 26 seconds: the speed promise is fragile. On a short task, the orchestration overhead can wipe out the advantage. On longer tasks that are genuinely decomposable, the gap should widen — but that's a hypothesis, not a measurement.

Another dry point: no accuracy advantage was measured. Parallel agents don't think better than the model in single-stream mode. They work at the same time, that's all. Any marketing reading that suggests otherwise would go beyond what the data shows today.

The real gain is inspection — and nobody's talking about it

Direct answer: the most concrete benefit of parallel agents isn't speed — it's being able to see the work.

In a classic single-stream response, the reasoning remains largely invisible: you get a final text, period. With parallel agents, Kingy.ai was able to observe 3 named agents, each openable in its own panel, with its sources and caveats. GenZ NewZ, which covered the test, reaches the same conclusion: the real value lies in this inspection surface that single-stream responses don't provide.

My take: this is the real strategic novelty. ChatGPT is shifting from an answer machine to a workspace where the process becomes auditable. For anyone who has ever needed to verify a model's claims — which is to say, everyone — being able to open each agent's work changes the posture. You no longer read an answer; you examine a case file.

This is exactly the discipline we describe in our article on AI agent red teaming: verifying what an agent actually did rather than trusting its final answer. OpenAI is industrializing access to this verification, even if it doesn't put it front and center in its communications.

Two drawbacks noted by Kingy.ai: the activity panel displayed "No outputs yet" even though all 3 agents had finished, and there is no per-run usage counter. At this stage, there's no way to measure what a parallel run actually costs in quota.

What OpenAI Isn't Saying (Yet)

Direct answer: three key questions remain without an official answer — how many agents, at what quota cost, and who decides how tasks get split.

The maximum number of concurrent agents is announced nowhere. GoldenEra also notes the absence of any quantified benchmark to back up the "faster" claim, the absence of dedicated pricing, and the absence of a list of compatible plugins. A product update without a single performance number: that's rare, and it says something.

The impact on quota is the most sensitive question. Running multiple agents mechanically consumes more compute, and therefore potentially more of your subscription. OpenAI publishes no figures, and the lack of a per-run counter — the bug flagged by Kingy.ai — makes any precise tracking impossible. iaenunminuto sums up the three open questions: how many agents are allowed, whether it consumes more quota, and when enterprise workspaces will get it.

Last gray area, and the most interesting in my view: who decides how the work gets split? Neither the release notes nor the X post specify whether it's the model that chooses to break the task down, or whether the user can impose or adjust that split. In the Kingy.ai test, the brief was split into 3 lanes and the agents appear with names — which suggests model-side splitting. As it stands, you're at the mercy of the orchestration: you observe it, you don't steer it.

Finally, the rollout to Team accounts remains an open question, as does a possible GPT-6.1 Astra Pro, as reported by TestingCatalog via Digg. Worth watching in the coming weeks.

Facing Orca and multi-stream, OpenAI is catching up more than it innovates

Direct answer: agent parallelism wasn't born on October 16 — it's been running in developer tooling for months. What's new with OpenAI is distribution.

On the open source side, Orca already orchestrates fleets of coding agents in parallel, with a community that validated the approach within months (our deep dive). The multi-stream concept — bringing parallel processing into the very heart of inference — has likewise been documented for a long time (Multi-stream LLMs: why the future of AI agents runs through parallel processing). A developer who has been orchestrating multiple agents from a terminal won't discover anything fundamentally new.

What OpenAI is doing, on the other hand, is putting this mechanism one click away for millions of users who will never open a terminal. Historically, that's how real changes spread: not when the technology exists, but when it disappears behind the interface. Multi-agent needed its "mainstream" moment — here it is.

A telling detail: as early as October 17, a separate update to Codex and ChatGPT Work allows subagents to reuse context, extends context cache availability, and adjusts routing to increase the cache hit rate, according to TokenPost. In other words: OpenAI is already engineering the costs of this parallelization. The race to scale began even before the feature was rolled out broadly.

And the competition isn't standing still. Claude Opus 4.7 (Adaptive) and Gemini 3 Pro Deep Think are pushing hard on the agentic front, while GPT-5.5 still dominates the segment's leaderboards. If you're torn between ecosystems, our comparisons Claude vs ChatGPT and ChatGPT vs Gemini remain the best entry points — and our article on collaboration between multiple AIs explains why orchestration often matters more than the underlying model.

How to use it without wasting your quota

Direct answer: reserve parallel runs for tasks that can genuinely be broken down into independent subtasks — not for your everyday questions.

Good use cases: documentary research with several independent angles, competitive monitoring handled player by player, proofreading a long document section by section, or a brief split into lanes as in Kingy.ai's protocol. In all these cases, each agent produces a deliverable you can verify separately — that's where inspection really comes into its own.

Bad cases: any sequential task where step N+1 depends on the result of step N. Parallelism then brings nothing but noise, waiting, and consumption. And for a simple question, launching three agents is using a sledgehammer to crack a nut: you may end up paying several times over for an answer the single-stream mode delivers just as well.

Since no counter exists, impose manual discipline on yourself. A parallel run now and then, not systematically. An eye on your consumption after each use. And the habit of opening each agent panel rather than reading only the final synthesis — that's precisely where the value lies.

Checklist before launching a parallel run

• Can the task be broken down into genuinely independent subtasks?
• Will each agent produce a verifiable deliverable, with sources and caveats?
• Do you have enough quota headroom for a run that may consume several times more?
• Are you ready to inspect each panel, not just the overall conclusion?

If even one answer is no, stick to single-stream. You'll lose less time and less quota.

❌ Common Mistakes

Mistake 1: Expecting a speed gain proportional to the number of agents

Three agents don't make things three times faster. Kingy.ai measured a 9% average gain, with a first run slower than the sequential baseline. Orchestration has a cost, and it's paid in seconds. The solution: use parallelism for coverage and inspection, not for raw speed.

Mistake 2: Confusing parallel with smarter

No accuracy gain was measured in the independent test. Parallelism changes throughput and work visibility, not the quality of reasoning. The solution: judge each agent individually, panel by panel, instead of assuming that multiplying agents equals validation.

Mistake 3: Ignoring the quota cost

Multiple agents mean more compute, but there is no per-run counter and OpenAI publishes no figures. You can exhaust your quota without ever knowing what a specific run cost. The solution: track your consumption manually and reserve parallel runs for tasks that truly deserve them.

Mistake 4: Assuming the feature is available everywhere

Eligible Pro users first, managed workspaces depending on settings and rollout eligibility — and the precedent of the "dots" agents, excluded from the EEA, Switzerland, and the UK at launch. The solution: check your eligibility and region before rebuilding a workflow around this feature.

❓ Frequently Asked Questions

How many agents can work in parallel?

Nobody knows precisely. OpenAI hasn't published any limit, and the Kingy.ai test used three without hitting any block. The number of agents consumed per run isn't documented either. In practice, absent a usage counter, stick to modest splits — two to three agents — while waiting for official specifications.

Does it consume more quota?

Mechanically yes: multiple agents mean more compute. But no figures have been published and no per-run counter exists for now. The cache and routing adjustments deployed on October 17 suggest OpenAI is actively optimizing this cost. Monitor your consumption manually after each parallel run.

Who decides how the task is split?

OpenAI doesn't specify. In the independent test, a fixed brief was split into 3 lanes and the agents appear named in the interface, which suggests the split happens on the model side. No manual control over the split is documented at this stage: you observe the orchestration, you don't drive it.

Is it available on Team or in enterprise workspaces?

The feature is provided to the Pro 100, Pro 200, Business, and Enterprise plans, but in managed workspaces, availability depends on workspace settings and rollout eligibility. The extension to Team accounts remains an open question. In enterprise, the decision rests with your administrator.

What's the concrete benefit over a standard response?

Inspection. Each agent is named, visible during execution, and its result opens in its own panel with sources and caveats. It's a verification surface that single-stream responses don't offer — valuable for research and monitoring, incidental for a simple question.

What exactly is GPT-6 Astra Pro?

The high-end model of the GPT-6 Astra family, reserved for Pro plans. The release notes call it "GPT-6 Pro", the X post "GPT-6 Astra Pro": same update, two names. The base model scored 64.6% on Terminal-Bench, with notable results in science and on ARC-AGI-3.

✅ Conclusion

By putting parallel agents in the chat rather than in a terminal, OpenAI has shifted ChatGPT from answering to working as a workspace — modest speed gains (9% measured), but a real breakthrough when it comes to inspecting the AI's work. To explore open source and local alternatives, our selection of the best AI agents awaits you.