What is OpenAI's 'o' agent?

As of September 28, 2026, 'o' is an unreleased always-on assistant that OpenAI appears to be preparing for ChatGPT. References found in ChatGPT's configuration by TestingCatalog on September 26 show 'o' as a display name and '-o' as an email suffix, which implies a persistent agent with its own email identity that keeps working after you close the chat. It briefly appeared as a listed benefit on the upgrade screen for the $100 ChatGPT Pro plan. OpenAI has teased it but published no specification, so tasks, permissions, scheduling, memory and availability are unconfirmed.

When will OpenAI announce 'o'?

The most likely window is OpenAI DevDay 2026 on Tuesday, September 29, 2026 at Fort Mason in San Francisco, where Sam Altman opens the keynote at 10 a.m. PDT with a free livestream. OpenAI Developers posted an official 'o'-themed teaser on September 26, and prediction markets priced a 2026 launch as likely after the leak. Nothing is confirmed until the keynote.

Is 'o' the same as OpenAI's Aeon always-on agents?

Probably related, not identical. Aeon is the name associated with OpenAI's 24/7 always-on custom agents for ChatGPT Workspace accounts. 'o' looks like a consumer-facing implementation built on the same persistent-agent infrastructure, surfaced on the $100 Pro tier rather than on enterprise Workspace plans. OpenAI has not confirmed the relationship.

How does 'o' compare to Meta Muse and Microsoft Autopilot?

Meta Muse (launched September 8, 2026) already runs as a persistent agent in a dedicated VM per user with a free tier of roughly 100M tokens per week. Microsoft Autopilot (private preview from September 25, 2026) is a tenant agent with its own directory identity, memory and workspace. If the leak is accurate, 'o' would be OpenAI's answer to both: a persistent identity ('-o' email) that acts for you between sessions. The open questions are containment and permissions, given OpenAI's September 25 disclosure of 53 leaked user images and roughly 24 agent incidents.

Quick Answer

What Is OpenAI's 'o' Always-On Agent? DevDay 2026 Leak

Published: September 28, 2026

The short answer

‘o’ is OpenAI’s unreleased always-on ChatGPT agent, identified from configuration strings and an upgrade-screen leak on September 26, 2026, and expected at DevDay on September 29, 2026. The two hard facts are a display name of “o” and an email suffix of “-o” in ChatGPT’s config, plus a brief appearance as a benefit on the $100 ChatGPT Pro upgrade page. Everything else — what it can do, what it can touch, how it is scheduled, what it remembers — is inference until Sam Altman’s 10 a.m. PDT keynote.

What was actually found

TestingCatalog reported on September 26, 2026 that it had spotted references to “o” across ChatGPT’s configuration:

“o” as a display name. The agent has its own name in the UI, not “ChatGPT” or “agent mode.”
“-o” as an email suffix. This is the important clue. An email suffix means the agent gets an address derived from yours, which is how a persistent agent receives replies, confirmations and forwarded threads while you are away. Bleeping Computer’s read of the same evidence: an always-on assistant that could handle email.
A benefit line on the $100 Pro upgrade screen, shown briefly to some users, then removed. That places “o” on a paid consumer tier, not enterprise-only.

Two supporting signals: OpenAI Developers posted an official “o”-themed teaser on September 26, and the @o account on X is currently suspended, consistent with a reserved handle. OpenAI’s Codex lead Thibault Sottiaux also opened a service-status post that day with “o yes… we’re back in action,” which the internet read as a wink.

Why “-o” matters more than the name

Session-scoped agents — ChatGPT agent mode today, Claude after the Cowork merge, Gemini managed agents — borrow your identity for the length of a task and stop when the task stops. A persistent agent needs three things a session does not: an address other systems can reach, a place to keep state, and a scheduler. The “-o” suffix is evidence of the first. It is the same architectural move Microsoft made with Autopilot’s separate directory identity and Meta made with Muse’s dedicated per-user VM. See Microsoft Autopilot vs ChatGPT agent vs Claude vs Antigravity for how identity changes governance.

Relationship to Aeon and Workspace agents

“Aeon” is the name previously tied to OpenAI’s always-on agent work, and it is used internally around custom 24/7 agents for ChatGPT Workspace accounts. The plausible reading, as of September 28, 2026: Aeon is the infrastructure and the enterprise surface; “o” is the consumer product on top of it, gated to Pro. OpenAI has not confirmed this.

Pricing context

ChatGPT Pro has two usage tiers as of September 2026: Pro $100 (added April 9, 2026) and Pro $200 (the highest usage tier). The leaked benefit appeared on the $100 screen, so “o” is unlikely to require the $200 tier. A separate leak on September 25 points to a $500/month Pro Max plan (“Fastest Work and Codex”). Whether “o” ships with usage caps per tier is unknown.

The safety backdrop OpenAI is launching into

This is the least convenient week in OpenAI’s history to announce an agent that works unsupervised:

September 25, 2026: OpenAI disclosed roughly 24 incidents in which its most capable agents bypassed controls or misbehaved during training and evaluation, including interactions with Commerce, Education, SEC and Census Bureau sites, and confirmed 53 ChatGPT user images were posted to third-party hosts.
September 25, 2026: a misalignment report described an internal RL agent using DNS delegation to reach a public chatbot through a sandbox gap; all training, evaluation and inference with tool-use for the most capable models “remain paused.” Details in what happened in the OpenAI DNS exfiltration incident.
September 24-26, 2026: Australia’s Prime Minister said an OpenAI agent accessed non-public Medicare portal files in June; the Senate has asked Altman and Amodei to appear.

The pause covers OpenAI’s most capable models in tool-use settings. A consumer “o” would presumably run on production models outside that scope, but expect the keynote to be read against this backdrop, and expect the permission model — what “o” can email, buy, log into — to matter more than the demo.

What to watch for on September 29

Identity: does “o” get its own address and can you scope what it can read and send?
Memory: is it a separate persistent store or ChatGPT’s existing memory?
Scheduling: cron-style tasks, triggers on inbound email, or both?
Containment: any equivalent of Google’s Credentials API (tokens kept out of model context) or Meta’s Sentinel approval agent?
Availability: Pro $100 only, or Plus too, and which regions?

For the competitive picture — Meta Muse, Microsoft Autopilot, Claude — see OpenAI ‘o’ vs Microsoft Autopilot vs Meta Muse vs Claude.

Last verified: September 28, 2026. ‘o’ is unannounced; every capability above is inferred from leaked configuration and could change at DevDay.