OpenAI Spaces is the name of an unannounced ChatGPT feature that surfaced in the early hours of 29 September 2026, the day of OpenAI’s DevDay conference in San Francisco. TestingCatalog, which tracks unreleased AI features, reported that “a new announcement introducing ‘Spaces’ was added recently” to OpenAI’s code, and suggested Spaces “could be collaborative environments where users and agents work together on generated artifacts.”
That description is speculation, and TestingCatalog labelled it as such. But it fits a pattern that OpenAI has followed for two years: Canvas in 2024, shared projects and group chats in 2025, and workspace agents in April 2026. Each step moved ChatGPT from a private chat towards a shared place where several people, and now several AI agents, work on the same thing.
This article sets out what actually leaked, which DevDay sessions point in the same direction, how OpenAI Spaces would compare with rival shared workspaces, and what businesses should decide before people and agents start editing the same documents.
Table of contents
- What Leaked About OpenAI Spaces
- The DevDay Sessions That Point to OpenAI Spaces
- What OpenAI’s Codex Team Said About Collaboration
- How OpenAI Spaces Would Extend Canvas, Projects and Group Chats
- What a Human and Agent Workspace Needs
- How OpenAI Spaces Compares With Rival Shared Workspaces
- How to Read the OpenAI Spaces Leak
- What OpenAI Spaces Could Mean for Businesses
- When OpenAI Spaces Could Be Announced
- OpenAI Spaces FAQs
- References
What Leaked About OpenAI Spaces
The report came from TestingCatalog on X at 00:47 UTC on 29 September, about 16 hours before the DevDay keynote. The post was short. “OpenAI is working on ‘Spaces’,” it began, and “a new announcement introducing ‘Spaces’ was added recently.” An image of the announcement accompanied it, but with a note: “Announcement strings redacted.”
What the announcement string tells us
An announcement string is the text of an in-app notice that tells users about a new feature. Its presence suggests OpenAI Spaces had reached the stage of launch copy, not just an internal codename. It does not say what the feature does, who gets it or when it ships, and because TestingCatalog redacted the wording, the public cannot read the copy itself.
What TestingCatalog speculated
The rest of the post was clearly marked as a guess. Spaces “could be collaborative environments where users and agents work together on generated artifacts (speculation).” TestingCatalog also linked it to two DevDay sessions: one about a “Codex run that spent more than 1,000 hours” and one on “Multiplayer Codex”, which “might be closely connected (speculation).” By early afternoon UK time the post had been viewed more than 48,000 times.
The DevDay Sessions That Point to OpenAI Spaces
Hours earlier, at 23:30 UTC on 28 September, TestingCatalog had published highlights from DevDay’s session list, noting that “this agenda is not public yet.” Several sessions describe exactly the kind of shared, long-running agent work that OpenAI Spaces would need to host.
| Session | Description (as posted by TestingCatalog) | Link to Spaces |
|---|---|---|
| What Will You Build With Coding Agents That Don’t Quit? | A Codex run that “spent more than 1,000 hours keeping builds and tests passing in a large shared codebase” | Long-running agents working on shared material |
| From Single Player to Multiplayer with Codex | “Follow a 2am outage from the first Slack messages to a reviewed fix and an update for the team” | Several people and agents on one problem |
| What Comes After Loops? | “The evolution from simple loops to collaborating agents” | Agents that work with other agents |
| Codex Has Left The Laptop | Letting Codex “tackle entire problems and goals on its own” | Work that continues outside one person’s device |
| Claw Labs | Peter Steinberger on “how agents can change the way open-source teams build” | Teams, not individuals, as the unit of work |
| Codex Game Studio | “What makes a game easy for humans but hard for AI?” | Weak; a hands-on build session |
Coding agents that do not quit
The most striking session describes “an internal OpenAI agent” that “had goals, subagents, and persistent plans that helped it investigate failures and coordinate fixes” over more than 1,000 hours. An agent that runs for weeks needs somewhere to keep its plans, its outputs and its history where people can see and correct them. That is one plausible job for OpenAI Spaces.
From single player to multiplayer
The “Multiplayer” session follows an incident from Slack messages to a reviewed fix and a team update. The phrase “reviewed fix” matters: it implies a person approving what an agent produced. A shared workspace where agent output waits for human review is the natural home for that loop.
Platform tools inside Codex
TestingCatalog’s agenda post also said OpenAI “will build Platform capabilities inside Codex, enabling users to develop plugins, agents, and apps right from a dedicated DevDay page.” If developers can build agents inside Codex, OpenAI Spaces could be where those agents meet the people they work for.
What OpenAI's Codex Team Said About Collaboration
There is one on-the-record hint from OpenAI itself. On The Pragmatic Engineer podcast, released on 9 September, Thibault Sottiaux, who host Gergely Orosz noted “started building” Codex, said: “We have a couple of other things that are just really, very helpful for team productivity and team collaboration that we haven’t released yet, but are going to come, like some of it at DevDay.” Tibor Blaho, who tracks OpenAI’s apps, highlighted the quote on X the next day with its timestamp, 42:59.
The ChatGPT and Codex merge
The same podcast covered “The Merge: ChatGPT + Codex”. Blaho reported on 25 September that OpenAI had been rolling out a new version of the ChatGPT web app, “referred to as ‘Web Merge’ in the code (combining ChatGPT and Codex).” A single app for chat and agent work would make a shared space for both far easier to build.
What the quote does not confirm
Sottiaux did not name OpenAI Spaces, and “team collaboration” could describe many features. The quote shows that collaboration tools were planned for DevDay. It does not show that Spaces is one of them, or what form it takes.
How OpenAI Spaces Would Extend Canvas, Projects and Group Chats
If the speculation is right, OpenAI Spaces would be the fifth step in a clear sequence. Each earlier launch widened who could work together inside ChatGPT, and the last one brought agents into the group.
| Date | Launch | Who works together |
|---|---|---|
| 3 October 2024 | Canvas | One person and ChatGPT, editing the same document or code |
| 13 December 2024 | Projects | One person, with chats, files and instructions grouped together |
| 24 October 2025 | Shared projects for Free, Plus and Pro users | Several people sharing chats, files and instructions |
| 13 November 2025 | Group chats (pilot, global from 20 November) | Several people and ChatGPT in one conversation |
| 22 April 2026 | Workspace agents | Teams and shared, always-on agents |
| 29 September 2026 (leak) | OpenAI Spaces | Possibly people and agents on shared artefacts |
Canvas: one document, two editors
Canvas, launched on 3 October 2024, was OpenAI’s first attempt to make ChatGPT a co-editor rather than a chat partner. A person and the model worked on the same document or piece of code side by side. It was collaboration, but only between one human and one AI.
Projects, then shared projects
Projects arrived on 13 December 2024 as a way to group chats, files and instructions. Sharing followed: shared projects reached Free, Plus and Pro users on 24 October 2025, turning a private folder into a team space. OpenAI Spaces would plausibly go one step further by letting agents, not just people, contribute to that shared material.
Group chats and workspace agents
Group chats brought several people into one conversation with ChatGPT, starting as a pilot in Japan, New Zealand, South Korea and Taiwan on 13 November 2025 and expanding globally on 20 November. Then, on 22 April 2026, workspace agents let teams build shared, always-on agents with admin controls, approvals and scheduling. The missing piece is a common place where those agents and their human colleagues work on the same artefacts.
The gaps between launches have shortened and then settled into a steady rhythm. The chart counts the days between each milestone in the table above.
What a Human and Agent Workspace Needs
A shared document is easy. A shared document that several people and several AI agents can all change is not. Whatever OpenAI Spaces turns out to be, it will be judged on a short list of controls that decide whether a business can trust it.
| Requirement | Why it matters with agents | What to look for at launch |
|---|---|---|
| Attribution | People need to know whether a person or an agent made each change | Per-edit labels naming the agent and its owner |
| Version history | Agents make many small edits quickly | Full history with one-step rollback |
| Permissions | An agent should not see or change more than its owner can | Role-based access that applies to agents as members |
| Approval gates | Some actions, such as sending or publishing, need a human | Admin-set rules like those on workspace agents |
| Audit and export | Regulated teams must show who did what | Coverage by OpenAI’s Compliance API and exportable logs |
| Cost visibility | Long-running agents consume credits in the background | Usage shown per space and per agent |
Lessons from workspace agents
OpenAI already has a template for some of this. When workspace agents launched, TestingCatalog reported that admins “can determine which tools and actions an agent may use, require approval for steps like editing spreadsheets, sending emails, or creating calendar events, and monitor agent configurations and runs through the Compliance API.” OpenAI Spaces would need the same controls to apply inside a shared artefact, not just to an agent’s actions elsewhere.
The cybersecurity angle
A space that many people and agents can write to is also a new target. Prompt injection, where hidden instructions in a document steer an agent, becomes more dangerous when that agent can edit material others rely on. Cybersecurity teams will want to know how OpenAI Spaces isolates agents from untrusted content before they approve it.
How OpenAI Spaces Compares With Rival Shared Workspaces
OpenAI is not first to this idea. Several rivals already offer a shared place where people and AI work on the same material, although few yet treat agents as full members.
| Product | Launched | What people share | Role of AI |
|---|---|---|---|
| Microsoft Copilot Pages | September 2024 | Editable pages built from Copilot answers | Assistant that drafts content for the team to edit |
| Perplexity Spaces | October 2024 | Research threads, files and instructions | Search assistant scoped to the space |
| Anthropic Claude Projects | June 2024 | Project knowledge and chats | Assistant grounded in shared files |
| ChatGPT shared projects | October 2025 (all users) | Chats, files and instructions | Assistant grounded in shared files |
| OpenAI Spaces (leaked) | Not announced | Possibly generated artefacts | Possibly agents as co-workers alongside people |
Anthropic, Meta and the agent race
The competition is moving fast. Anthropic has been merging its chat and agent products, as we covered when it merged Claude Chat and Cowork. Smaller firms are experimenting too: Rene launched a multiplayer AI agent for iMessage calendar coordination in September. The difference OpenAI Spaces could bring is scale, because it would sit inside a product that already has hundreds of millions of users.
A crowded name
“Spaces” is also one of the most reused names in software. X Spaces is live audio, Hugging Face Spaces hosts machine learning demos, Google Chat has spaces for team conversations, DigitalOcean Spaces is object storage, and Perplexity Spaces is a research workspace. If OpenAI keeps the name, search results and support tickets will need care, and it would not be the first time a leaked name failed to survive launch.
How to Read the OpenAI Spaces Leak
Leaks about OpenAI have been unusually frequent in the run-up to DevDay, and not all of the coverage has been careful. One crypto news site published a story at 00:52 UTC on 29 September saying OpenAI “showcased” Spaces at DevDay, hours before the keynote began. The same article conceded that “no product named ‘Spaces’ has been formally confirmed.”
What is solid and what is not
The evidence for OpenAI Spaces is a single announcement string, spotted by a reliable tracker and shown only in redacted form. That is enough to say a feature called Spaces existed in OpenAI’s code on 29 September. Everything about its purpose, including the human-and-agent reading, is inference from the name and from the DevDay agenda.
Names change before launch
OpenAI’s own recent history shows how often that happens. Its always-on agent builder appeared in TestingCatalog’s reporting under the codename “Hermes” in April, and shipped a day after the fullest report as “workspace agents”. The consumer agent that leaked this week has already appeared as “Aeon”, then “o”, then “dots”, which we covered in our report on OpenAI Dots. OpenAI Spaces could launch under another name, or not at DevDay at all.
What OpenAI Spaces Could Mean for Businesses
Even unconfirmed, the direction is clear enough to plan for. If people and agents are going to share workspaces, the rules for that need to exist before the feature arrives, not after staff have started using it.
Decide where shared artefacts live
Work produced in an AI workspace can end up outside your document management system, retention policy and search. Decide now whether shared AI artefacts must be exported to your main systems, and who owns them when a project ends.
Treat agents as members with permissions
Give each agent an owner, a purpose and the minimum access it needs, and review that access like you would for a contractor. Our guide to AI employees and autonomous agents covers how to set those roles up, and good workflow automation design decides where a human must approve.
Plan for audit and rollback
Ask whether every agent edit is attributed and reversible, and whether logs reach your compliance tools. If they do not, keep agents on draft-only permissions in shared spaces until they do.
Prepare people, not just tools
The hardest part is cultural. Staff need to know when to trust an agent’s contribution and when to check it. Our piece on how AI agents are about to flood the workforce covers the change-management side.
When OpenAI Spaces Could Be Announced
DevDay 2026 takes place on 29 September at Fort Mason in San Francisco. OpenAI’s schedule lists the opening keynote at 10:00 a.m. Pacific Time, “featuring Sam Altman”, followed by breakout sessions from 11:15 a.m. to 3:30 p.m. and a closing session at 4:00 p.m. The keynote is livestreamed, and OpenAI teased “20+” new launches the evening before, according to TestingCatalog.
| Event (29 September 2026) | San Francisco (PDT) | UTC | London (BST) |
|---|---|---|---|
| Opening keynote | 10:00 | 17:00 | 18:00 |
| Breakouts begin | 11:15 | 18:15 | 19:15 |
| Breakouts end | 15:30 | 22:30 | 23:30 |
| Closing session | 16:00 | 23:00 | 00:00 (30 September) |
The London connection
OpenAI is also running “DevDay Exchanges” in Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo and Mexico City. UK developers who want to see OpenAI Spaces, if it appears, in a hands-on setting should watch for the London event.
How much attention the leaks drew
Interest in the pre-DevDay leaks was high. The chart compares the view counts of TestingCatalog’s three posts, as read at about 13:00 UTC on 29 September.
OpenAI Spaces FAQs
What is OpenAI Spaces?
OpenAI Spaces is an unannounced ChatGPT feature whose announcement text was spotted in OpenAI’s code on 29 September 2026. TestingCatalog speculated that it could be a collaborative environment where users and AI agents work together on generated artefacts.
Has OpenAI confirmed Spaces?
Not before the DevDay keynote on 29 September 2026. The only evidence is a redacted announcement string reported by TestingCatalog. DevDay may confirm, rename or omit it.
How is it different from shared projects?
Shared projects let several people share chats, files and instructions. If the speculation is right, OpenAI Spaces would add AI agents as active contributors to shared work, rather than a single assistant answering questions.
Is it connected to Multiplayer Codex?
Possibly. TestingCatalog linked Spaces to DevDay sessions on “Multiplayer Codex” and a 1,000-hour Codex run, but described the connection as speculation.
When will businesses get it?
OpenAI has not said. Workspace agents launched first on Business, Enterprise, Edu and Teachers plans, so a business-first rollout is plausible, but it is not confirmed.
References
OpenAI is working on “Spaces” (TestingCatalog on X)
Key insights from upcoming DevDay sessions (TestingCatalog on X)
Thibault Sottiaux on team collaboration features at DevDay (Tibor Blaho on X)
Group chats in ChatGPT (OpenAI)
OpenAI launched 24/7, always-on Workspace Agents in ChatGPT (TestingCatalog)
OpenAI develops platform for always-on Agents on ChatGPT (TestingCatalog)
Microsoft 365 Copilot Wave 2: Pages, Python in Excel, and agents (Microsoft)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
