OpenAI adds multi-user group chats to ChatGPT

The interface shift: one conversation, many senders

Until now, a ChatGPT conversation had exactly two participants: one user and the model. Group chats break that assumption. You tap a people icon, share a link with one to twenty people, and everyone lands in the same thread alongside ChatGPT.

OpenAI made one careful choice at the entry point: adding someone to an existing chat doesn't expose that chat. It creates a copy as a new group chat, leaving the original private conversation untouched. That copy-on-invite behavior is a small detail with a large privacy consequence — it prevents accidental disclosure of an ongoing solo session.

Participants set up a short profile with name, username, and photo, and group chats live in their own labeled sidebar section. This is the first time ChatGPT has needed a concept of 'who is speaking' inside a thread at all.

Teaching a chatbot when not to talk

A one-to-one assistant responds to every message by default, because every message is addressed to it. In a group of people, that behavior would be intolerable. So OpenAI describes teaching ChatGPT 'new social behaviors': it follows the conversation and decides when to respond and when to stay quiet based on context.

It follows the flow of the conversation and decides when to respond and when to stay quiet based on the context of the group conversation. You can always mention "ChatGPT" in a message when you want it to respond.Montana Labs

The explicit @-mention fallback is the tell: turn-taking by inference is unreliable, so they kept a deterministic override. The model also gained the ability to react to messages with emojis and to reference participants' profile photos — for instance, using group members' photos in generated images. These are presence signals borrowed from human messaging apps, retrofitted onto an assistant that previously had no notion of a room.

Billing and model selection follow the person being answered

A shared thread creates a genuinely new problem: whose subscription applies when the participants are on different plans? OpenAI's answer is per-response attribution. Responses run on GPT-5.1 Auto, which picks a model based on the plan of the specific user ChatGPT is replying to — Free, Go, Plus, or Pro.

Rate limits follow the same rule. Messages between humans in the group don't count against anything; only ChatGPT's responses do, and each counts against the limit of the person it answered. This keeps a Pro user's quota from silently subsidizing an entire group, and it means the same group chat can produce different-quality responses depending on who prompted.

Memory is walled off, and minors reshape the whole room

Personal ChatGPT memory is not used in group chats, and the model creates no new memories from them. OpenAI says it's exploring more granular controls later, but for now the default is strict isolation — a sensible position when a single conversation may contain people who don't otherwise share data.

The most consequential safeguard is age-driven and applies to everyone. If anyone under 18 is in the chat, ChatGPT automatically reduces exposure to sensitive content for all participants, and guardians can disable group chats entirely through parental controls. The presence of one minor changes the model's behavior for the whole group — a policy that treats the room, not the individual, as the unit of safety.

What a shared assistant demands from product teams

The pilot launched in Japan, New Zealand, South Korea, and Taiwan on November 13, 2025, and by November 20 OpenAI expanded it to all logged-in Free, Go, Plus, and Pro users globally after positive early feedback — a fast graduation from regional test to worldwide rollout.

The lesson for anyone building assistant features is that moving from one user to many is not a UI addition; it changes the core contract. You suddenly need identity per message, a policy for when the model speaks, per-user billing attribution inside a shared context, and memory boundaries that hold across people who never agreed to share data. OpenAI answered each with an explicit rule rather than a default. Teams adding collaboration to an AI product should expect to make the same four decisions, and to make them before shipping, not after.