Always-on assistant is the phrase that gave OpenAI’s next product away. On Friday 25 September 2026, a ChatGPT user named Jake Boggs posted a screenshot of the “Get ChatGPT Pro” upgrade screen with an unfamiliar line circled in yellow: “o, your always-on assistant”. Within a day the same words had been found in ChatGPT’s own web code, a configuration entry had surfaced giving the product a display name and an email suffix, and OpenAI’s Thibault Sottiaux had opened two posts on X with a lowercase “o”.
OpenAI has announced nothing. But the evidence is unusually concrete, and it lands four days before OpenAI DevDay in San Francisco on Tuesday 29 September. We downloaded the ChatGPT code file ourselves. It contains the line, together with a note for translators that reads “Lowercase o is the assistant’s product name.” That makes the name close to certain, even if the launch date is not.
This article sets out what leaked and when, what the evidence says OpenAI’s always-on assistant would do, which ChatGPT Pro plans appear to include it, and the Codex “Persistent mode” work that looks like its engine. It compares o with Meta’s Muse, xAI’s Grok Bot and Anthropic’s Conway, weighs the risks of an agent that never clocks off, and ends with the decisions a business should make before switching one on. It builds on our coverage of OpenAI’s Managed Agents leak ahead of DevDay and the $500 ChatGPT Pro Max tier.
Table of contents
- What Leaked About OpenAI’s Always-On Assistant
- What the Always-On Assistant Would Be
- The Codex Code Behind OpenAI’s Always-On Assistant
- Which ChatGPT Plans Include the Always-On Assistant
- DevDay 2026 and the Always-On Assistant
- How the Always-On Assistant Compares With Muse, Grok Bot and Conway
- Before the Always-On Assistant: OpenAI’s Earlier Attempts
- The Risks of an Always-On Assistant
- What Businesses Should Decide Before Using an Always-On Assistant
- What We Still Don’t Know About the Always-On Assistant
- Frequently Asked Questions About OpenAI’s Always-On Assistant
- References
What Leaked About OpenAI's Always-On Assistant
The leak arrived in four pieces over about 21 hours. None of them is an announcement, but together they show the same product from four sides: the paywall, the code, the configuration and the people who build it.
The upgrade screen
Boggs posted the screenshot at 14:50 UTC on 25 September, asking whether he had “missed a product launch” or found “a dev day leak”. It shows a $100.00 price labelled “Standard” and a benefits list headed “Everything in Plus, and:”. The second item, between “Most capable frontier models” and “More Work and Codex usage”, is “o, your always-on assistant”. The rest of the list is familiar: maximum memory, 100 GB of file storage and early access to new features. By Saturday morning the post had passed 400,000 views.
The line in ChatGPT’s own code
The strongest evidence is first-party. A ChatGPT web bundle, last modified at 21:31:52 GMT on 25 September, holds the pricing-page strings for the iOS app. We fetched it and found the entry verbatim: the text “o, your always-on assistant”, with the description “Benefit shared by all Pro tiers on the iOS pricing page. Lowercase o is the assistant’s product name.” Descriptions like that are written for translators, which is why the name is spelled out so plainly.
The translations are already in place. The German bundle renders the always-on assistant as “o, dein Assistent rund um die Uhr”, and we found the same entry in the Japanese, Hindi and Korean bundles, each keeping the lowercase “o” untranslated. The X user Chetaslua, who first posted the code, counted 63 languages. We did not check all of them, but four out of four is enough to show the strings were prepared for a global release rather than a one-market test.
The configuration entry
TestingCatalog, which tracks unreleased features, posted at 10:02 UTC on 26 September that “an upcoming always-on assistant from OpenAI will be named ‘o'”. Its screenshot of a ChatGPT dynamic configuration shows a value block containing "email_suffix": "-o" and "display_name": "o". TestingCatalog reads the suffix as a sign that o “will support email handling”. That is a reasonable inference, but the entry does not show how the suffix is applied, so treat email as likely rather than documented.
The “o no” and “o yes” posts
Late on 25 September, Codex and ChatGPT Work suffered an outage. Sottiaux, whom Wired describes as OpenAI’s head of core products, posted “o no :(” at 23:08 UTC. An hour later he followed with “o yes… we’re back in action and we’ll reset usage limits for all paid users across codex and ChatGPT work”. The second post passed three million views. Some commentators and prediction-market traders read the phrasing as a tease. It may simply be a joke, and it is the weakest evidence in the set.
| Time (UTC) | Source | What it shows | Evidence type |
|---|---|---|---|
| 25 Sep, 14:50 | Jake Boggs on X | Pro upgrade screen listing “o, your always-on assistant” at $100 | Screenshot of a live screen |
| 25 Sep, 21:31 | ChatGPT web bundle | The string, “all Pro tiers”, and “Lowercase o is the assistant’s product name” | OpenAI code, verified by us |
| 25 Sep, 23:08 | Thibault Sottiaux on X | “o no :(” during a Codex and Work outage | Staff post, ambiguous |
| 26 Sep, 00:07 | Thibault Sottiaux on X | “o yes…” plus a usage-limit reset for paid users | Staff post, ambiguous |
| 26 Sep, 10:02 | TestingCatalog on X | Config with display name “o” and email suffix “-o” | Third-party screenshot |
| 26 Sep, 11:49 | Chetaslua on X | The same string in localised bundles, “all 3 Pro tiers”, 63 languages | Third-party code reading |
What the Always-On Assistant Would Be
Nothing OpenAI has published describes o’s features. What follows is built from the strings above, from the public Codex code covered in the next section, and from claims made by people who have read more of ChatGPT’s code than we have. Each claim is labelled with how solid it is.
A named assistant, not a setting
The name is the biggest clue. OpenAI’s existing agent features in ChatGPT are modes and tools that you pick for a task and that stop when the task ends. A product with a display name and its own email suffix sounds closer to a colleague: a persistent identity that can be addressed, can send and receive messages, and carries its work across days. “Always-on assistant” is also a different promise from “agent mode”. It describes availability, not a single job.
Email as the first sign of scope
The -o suffix is the only capability hint in the configuration itself. The entry does not say how it is used. One plausible reading is that the always-on assistant would get an address derived from the user’s, so that mail it sends or receives is marked as the assistant’s rather than the person’s. That would be the right design for accountability. It would also make o the first OpenAI product that acts in a user’s inbox as a separate, named sender.
A computer in the cloud
Tech commentator Mark Kretschmann wrote on 26 September that o “will be comparable to Grok Bot and Muse, with a computer in the cloud, which is always online”. We did not find that in the code we checked, so treat it as a claim. It is consistent with where OpenAI has been heading, though: Codex already runs tasks in cloud containers, and the Managed Agents leak described persistent environments for agents to work in.
Aeon and ChatGPT’s workspace agents
A rumoured OpenAI project called “Aeon” circulated in the days before the leak. TestingCatalog says Aeon is an internal name for the existing custom Agents implementation for ChatGPT Workspace accounts, not a separate product, and that OpenAI “will likely build a consumer-facing ‘o’ assistant on top of this feature”. If that is right, the always-on assistant is the consumer edition of machinery that businesses on ChatGPT workspaces already use.
Sixty-three languages
Chetaslua’s count of 63 localised versions has not been confirmed by OpenAI. Our spot check of four matched. Either way, the translation work shows OpenAI is preparing o for every market where the iOS Pro page appears, rather than for a small English-language beta.
The Codex Code Behind OpenAI's Always-On Assistant
The consumer name is new, but the engine that an always-on assistant needs has been visible for a month. OpenAI publishes the source of its Codex coding agent at openai/codex on GitHub, and a run of commits since late August adds exactly the pieces an agent needs to keep working after the conversation ends.
A “Persistent” reasoning effort
On 26 August, a commit titled “Support persistent reasoning effort” added persistent to Codex’s reasoning-effort settings, the menu where users choose how much computing time a model spends on a task. The terminal interface shows it as “Persistent”. Wired, which first reported the work on 27 August, found that in this mode Codex will “continue working until put to sleep”, where the current settings stop after minutes or hours even if a task is unfinished.
The proactivity instructions
A second commit on 27 August added a prompt file, persistent_mode.md, that tells the agent how to behave once its main task is done. It reads like a job description for an always-on assistant: “look for useful follow-ups that directly support the completed work”, “Use past user instructions and your knowledge of the user to prioritize follow-ups, not to infer new authorization”, and treat a follow-up “as active ongoing work across sleeps and automatic continuations until the outcome is established”.
The same file sets limits. “Persistence does not broaden that scope,” it says, and any follow-up that needs “new authority, materially expand scope, or make external state changes not already authorized” must be proposed to the user and approved first. It even sets a default rhythm for checking back: “often 1–3 minutes for active near-term work”, backing off when progress is slower.
Clocks, sleep and an “always_on” switch
Two more commits fill in the plumbing. One, on 26 August, switches on a current-time reminder and an interruptible clock.sleep tool for persistent turns, so the agent knows what time it is and can wait without ending. Another, on 28 August, adds a sleep_tool setting with two modes, model_driven and always_on. Together they let an agent pause, wake and resume on its own schedule, which is the core of any always-on assistant.
Message boards for agents
On 25 September, the same day the upgrade screen leaked, OpenAI merged a commit adding “an HTTP client for remote agent message boards”. It lets agents create boards, register members, post and receive live notifications, with previews capped at 150 characters. Commentators read it as a shared noticeboard where several agents can swap findings. That matters for o because a long-running assistant would need somewhere to hand work to, and collect it from, other agents.
What OpenAI said in August
When Wired asked about Persistent mode, an OpenAI spokesperson confirmed the company was testing it but said there were no immediate plans to launch it. Sottiaux added that “OpenAI is a very bottom-up culture and many different things are explored on the open source repo which is a bit of our shared playground.” A month later, a named always-on assistant is in the pricing strings. Plans evidently moved.
| Codex change | Date (2026) | What it adds | Why an always-on assistant needs it |
|---|---|---|---|
| #40799 | 26 Aug | “Persistent” reasoning effort | Keeps working instead of stopping at a time limit |
| #40942 | 26 Aug | Clock reminder and interruptible sleep for persistent turns | Waits and wakes without losing the task |
| #41050 | 27 Aug | Proactivity and follow-up instructions | Decides what to do next, within the user’s authority |
| #41243 | 28 Aug | Sleep tool with model_driven and always_on modes | Lets the product force the sleep tool on |
| #48100 | 25 Sep | Client for remote agent message boards | Shares findings with other agents between sessions |
Which ChatGPT Plans Include the Always-On Assistant
The code answers this more clearly than anything else. The bundle describes o as a “Benefit shared by all Pro tiers”, alongside “Most capable frontier models”, “100 GB file storage” and “Early access to new features”. It also names the three tiers through their short descriptions: “Standard” for the Pro Lite tier, “More usage” for Pro, and “Max usage” for Pro Max. Plus, at $20 a month, lists “ChatGPT Work assistant” and agent features, but not o.
So the always-on assistant looks like a Pro-only benefit, available from the cheapest Pro tier upwards. The tiers then differ on speed and volume: “More Work and Codex usage” for Pro Lite, “Faster Work and Codex” for Pro and “Fastest Work and Codex” for Pro Max. Prices come from different sources. The $100 figure is on Boggs’s screenshot, OpenAI’s Help Center describes the $200 plan as Pro 20x, and the $500 Pro Max price was reported this week but is not yet on a public OpenAI page.
| Plan | Monthly price | Label in the code | Work and Codex line | Always-on assistant o |
|---|---|---|---|---|
| Plus | $20 | Not a Pro tier | Work agent and Codex agent | Not listed |
| Pro Lite | $100 (screenshot) | “Standard” | “More Work and Codex usage” | Included |
| Pro | $200 (Help Center) | “More usage” | “Faster Work and Codex” | Included |
| Pro Max | $500 (reported) | “Max usage” | “Fastest Work and Codex” | Included |
Why the always-on assistant starts at Pro
Compute is the obvious reason. Wired described Persistent mode as one of OpenAI’s most computationally intensive settings, and an assistant that checks back every few minutes for days burns far more tokens than one that answers and stops. OpenAI has also been short of capacity: earlier this month it paused new ChatGPT Pro subscriptions because of demand for Astra. Gating an always-on assistant behind a $100 floor limits how many people can switch one on at launch.
What “all Pro tiers” does and does not tell us
“All Pro tiers” means the feature is present on every Pro plan. It does not mean the allowance is the same. The Work and Codex lines scale from “More” to “Fastest”, and it would be surprising if o’s running time or task volume did not scale with them. Pricing for heavy use, and whether o draws on the same usage limits Sottiaux reset on Friday night, are open questions.
DevDay 2026 and the Always-On Assistant
The timing is not subtle. OpenAI’s developer conference, DevDay, takes place in San Francisco on Tuesday 29 September, and TestingCatalog expects OpenAI to say more about the always-on assistant there. That would fit a pattern: consumer features increasingly launch alongside the developer tools they are built from.
What is confirmed about DevDay
OpenAI’s event page confirms the date and city, and TestingCatalog reports that Sam Altman will give the opening keynote. Earlier this month, TestingCatalog reported that OpenAI was preparing Managed Agents and agent environments for the event. Those are developer products, but an always-on assistant is the obvious consumer showcase for them.
What else to watch for
Commentators expect DevDay to include a “fast mode” served from Cerebras hardware, reportedly tied to the $500 plan, and the shared message boards for agents that appeared in Codex on Friday. Neither is confirmed. If o is announced, look for four details the leak does not settle: where it runs, how it handles email, what it may do without approval, and which countries get it first.
What the prediction markets did
Octagon, which tracks prediction markets, reported that Kalshi’s contract on OpenAI releasing a personal AI agent before 1 January 2027 jumped from 24% to 78% after the leak, and the contract for a release before 1 November reached 80%. Those are bets, not facts, but they show how seriously traders took the always-on assistant evidence.
How the Always-On Assistant Compares With Muse, Grok Bot and Conway
OpenAI is not first. Its rivals have spent the summer launching or testing agents that run in the background, and TestingCatalog’s own verdict was that “Meta set a super high standard with its Muse assistant. OpenAI will need to cross a very high bar.”
| Product | Company | Status on 26 September 2026 | How it is described |
|---|---|---|---|
| o | OpenAI | Unannounced, present in ChatGPT code | “your always-on assistant” for all Pro tiers |
| Muse | Meta | Launched 8 September | Personal agent that can make suggestions unprompted |
| Grok Bot | xAI | Launched 11 August | Desktop app with named AI teammates doing background work |
| Conway | Anthropic | Reported in development in July | Always-on agent for Claude on iOS |
Meta’s Muse
Muse is the benchmark. Meta launched it on 8 September as a personal agent that, in Meta’s words, can make suggestions unprompted. OfficeChai describes it as running on a secure virtual machine and handling travel booking and shopping through WhatsApp. Its early uptake has been strong, as we covered in Meta’s AI Tamagotchi bet, but so have its problems, including a bug-bounty vulnerability that forced a stronger safety warning.
xAI’s Grok Bot
Grok Bot launched on 11 August as a desktop app where users create named AI “teammates” that run tasks such as inbox management and code review in the background. It gained a voice mode in September, which we covered alongside Anthropic’s Claude Projects redesign. Kimmonismus, a widely followed AI commentator, called o “essentially a GrokBot alternative”.
Anthropic’s Conway
Anthropic’s answer, codenamed Conway, was reported in July as an always-on agent for Claude on iOS, as described in our Conway AI agent coverage. Google has its own proactive agent effort in Gemini Spark. Every major lab is converging on the same idea: an assistant that works while you do not.
Where o could stand out
OpenAI’s advantages are reach and plumbing. ChatGPT already has more than a billion weekly users, the Codex code gives o a tested engine for long tasks, and a dedicated email identity would be a clean answer to the question of who sent a message. Its disadvantage is timing: Meta and xAI have months of real-world feedback that OpenAI does not.
Before the Always-On Assistant: OpenAI's Earlier Attempts
OpenAI has tried to make ChatGPT proactive before, and none of those attempts became a habit for most users. That history explains both why o exists and why it might struggle.
Pulse
In September 2025, OpenAI launched Pulse, a feature that prepared a personalised morning briefing while users slept. Wired reported that OpenAI sunset Pulse earlier this summer, and called Persistent mode “a considerably more ambitious version of the same bet”. An always-on assistant is Pulse with the ability to act, not just to summarise.
Operator, agent mode and ChatGPT Work
Operator, OpenAI’s first web-browsing agent, launched in January 2025 and was folded into ChatGPT agent mode that July. This year OpenAI added ChatGPT Work for longer autonomous tasks. Each step let the model do more, but each still waited for the user to start it. The always-on assistant would be the first to start work on its own initiative.
Why OpenAI keeps trying
Sam Altman has been clear about the goal. “There’s like a single product which is: I need to ask the AI something,” he said on David Senra’s podcast, as quoted by Wired. “Eventually, maybe the AI should proactively offer me things.” Wired adds that OpenAI hopes proactive features will lift use of its most advanced models, which only a fraction of ChatGPT users touch today.
The Risks of an Always-On Assistant
An assistant that runs for days, knows a lot about you and can send email is useful for the same reasons it is risky. OpenAI’s own disclosures make the case better than any critic could.
Persistence and alignment
Wired reported that OpenAI’s recent Hugging Face hacking incident was driven mainly by an internal research model trained to be highly persistent, which OpenAI has since taken offline. OpenAI also said its agents, faced with impossible tasks, tried to probe and compromise the sandboxes they ran in, and that forthcoming models including Astra were trained to enable persistent agents. We looked at the wider pattern in OpenAI’s five ways rogue agents are messing with the internet.
Guardrails written into the prompt
The persistent_mode.md file shows OpenAI knows this. It tells the agent that persistence “does not broaden” its scope and that new authority or external changes need approval. Those are sensible rules. But they are instructions to a model, not hard permissions enforced by the system, and an always-on assistant will be judged on whether the two line up.
Email identity and impersonation
If o gets its own email identity, that address becomes a target. Attackers already imitate trusted senders, and a message that appears to come from a colleague’s always-on assistant could be very persuasive. Organisations should treat assistant mailboxes as a cybersecurity asset from day one: monitored, rate-limited and never trusted by default for payment or credential requests.
Memory and privacy
The proactivity file tells the agent to use “knowledge of the user” to prioritise its work, and the Pro screen promises “Maximum memory”. An always-on assistant that remembers more and acts more will hold a detailed record of a person’s work and life. How long that memory lasts, who can see it and how it can be deleted will matter far more than it did for a chatbot.
Reliability and cost
Friday’s outage, and Sottiaux’s promise to reset usage limits afterwards, is a reminder that a service which never sleeps depends on infrastructure that sometimes does. Gizmodo reported that Codex users already see the agent stop “on its own at times—often before actually finishing its tasks.” For an always-on assistant, silent failure is worse than a visible error.
What Businesses Should Decide Before Using an Always-On Assistant
Pro is sold to individuals, and nothing in the leak says whether o will reach ChatGPT Business or Enterprise workspaces. But staff with personal Pro plans will try it on work tasks the week it launches. The decisions below are worth making now, as part of any AI strategy, rather than after the first incident.
| Decision | Why it matters | A sensible default |
|---|---|---|
| Personal plans on work data | Pro is a consumer plan outside company controls | No company data in personal assistants |
| Actions without approval | Persistent agents act between check-ins | Read-only by default, approval for anything external |
| Email sending | An assistant address can be spoofed or misused | Separate identity, no payment or credential requests |
| Memory and retention | Long-running context holds sensitive detail | Document retention, test deletion |
| Logging and review | Proactive work happens when nobody is watching | Weekly review of what the assistant did |
Set a policy for personal assistants
Most organisations already have rules for consumer chatbots. Extend them explicitly to always-on assistants, because the risk is different: a chatbot answers a question, while o would keep working with whatever it was given. Put the policy under your existing IT governance process, with a named owner and a review date.
Decide what the always-on assistant may do alone
The proactivity file already separates safe, “non-mutating” follow-ups from actions that need approval. Mirror that split in policy. Reading, drafting and summarising can run unattended. Sending, buying, deleting and changing anything outside the user’s own workspace should always wait for a person.
Plan for a workforce of agents
An always-on assistant is one agent working for one person. The same questions apply at scale when teams deploy AI employees and autonomous agents across a business: identity, permissions, audit trails and cost controls. Building those controls once, for every agent, is cheaper than retrofitting them product by product.
What We Still Don't Know About the Always-On Assistant
The leak answers the naming question and the plan question. It leaves most of the product questions open. The table grades each claim by the strength of its source.
| Claim | Status | Source |
|---|---|---|
| The product is called “o” | Confirmed in OpenAI code | ChatGPT web bundle |
| Included in all Pro tiers | Confirmed in OpenAI code | iOS pricing strings |
| Cheapest tier costs $100 | Seen on a live screen | Boggs screenshot |
| Email suffix “-o” | Seen in a config screenshot | TestingCatalog |
| Runs on its own cloud computer | Third-party claim | Mark Kretschmann |
| 63 languages | Third-party count, 4 checked by us | Chetaslua |
| Built on Codex Persistent mode | Our inference from public code | openai/codex commits |
| Announced at DevDay on 29 September | Expected, not confirmed | TestingCatalog, Kretschmann |
| Available to Business and Enterprise | Unknown | No source |
Frequently Asked Questions About OpenAI's Always-On Assistant
What is OpenAI’s “o”?
“o” is the product name of an always-on assistant that OpenAI appears to be preparing for ChatGPT. ChatGPT’s own code lists it as a benefit of every Pro plan and says “Lowercase o is the assistant’s product name.” OpenAI has not announced it or described its features.
Is the always-on assistant confirmed?
The name and the Pro placement are confirmed by strings in ChatGPT’s code, which we checked directly. The features and the launch date are not. Until OpenAI publishes a launch post or a Help Center article, everything beyond the name should be treated as reported.
Which ChatGPT plan would I need?
Any Pro plan, according to the code. That means Pro Lite, shown at $100 a month on the leaked screen, Pro at $200 and the reported $500 Pro Max. The $20 Plus plan does not list o.
When will the always-on assistant launch?
OpenAI DevDay on Tuesday 29 September is the most likely moment for an announcement, and prediction markets priced a launch this year as likely after the leak. OpenAI often announces features before rolling them out to everyone, so availability could lag.
Is o the same as Codex Persistent mode?
Not exactly. Persistent mode is a setting in Codex, OpenAI’s coding agent, that keeps a task running until it is “put to sleep”. o appears to be a consumer product that would use the same kind of engine for everyday work, with its own name and email identity.
Should I let an always-on assistant handle work email?
Not without rules. Keep company data out of personal plans, require approval before the assistant sends anything on your behalf, and treat its mailbox as a security asset. Revisit the decision once OpenAI publishes how o’s permissions and data retention actually work.
References
Jake Boggs on X: ChatGPT Pro upgrade screen listing o
ChatGPT web bundle carrying the o pricing strings
TestingCatalog on X: OpenAI’s always-on assistant will be named o
TestingCatalog: OpenAI to announce o always-on agent during DevDay
Thibault Sottiaux on X: o yes, usage limits reset
Chetaslua on X: OpenAI o is in chatgpt.com’s code
OfficeChai: OpenAI could soon launch o, an always-on AI assistant
Wired: OpenAI is developing a persistent AI agent
openai/codex #40799: Support persistent reasoning effort
openai/codex: persistent_mode.md proactivity instructions
openai/codex #41243: Configurable gating for the sleep tool
openai/codex #48100: HTTP client for remote agent message boards
OpenAI Help Center: About ChatGPT Pro tiers
Octagon: OpenAI o agent leaks fuel bets on an imminent personal agent
Gizmodo: Nevertheless, OpenAI persists with new always-on agent
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
