One Letter, Always On: OpenAI's 'o' Agent Leaks Days Before DevDay

Four days before OpenAI’s DevDay in San Francisco, the company’s own shipping code appears to have spoiled the keynote. A lowercase letter — o — has surfaced across ChatGPT’s pricing screens, web bundles, and configuration entries, and every piece of evidence describes the same thing: an always-on assistant that keeps working after you close the chat, with its own email identity, sold as a benefit across all ChatGPT Pro tiers.

OpenAI has announced nothing. But the evidence is unusually concrete for a leak, because most of it is first-party: strings written by OpenAI’s own engineers, in OpenAI’s own product, with a translator’s note that spells the product name out in plain English.

How the leak unfolded, in four pieces

The first sign came on Friday, 25 September 2026, at 14:50 UTC, when 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.” The screen showed a $100.00 plan labelled “Standard,” with a benefits list reading “Everything in Plus, and:” — frontier models, more Work and Codex usage, maximum memory, 100 GB of file storage, early access to new features, and, sitting in the second position, the mysterious lowercase o. By Saturday morning the post had passed 400,000 views.

Hours later, the strongest evidence landed. A ChatGPT web bundle, timestamped 21:31:52 GMT on 25 September, was found to contain the pricing-page strings for the iOS app — verbatim: “o, your always-on assistant,” with an internal description that reads “Benefit shared by all Pro tiers on the iOS pricing page. Lowercase o is the assistant’s product name.” Those descriptions are written for translators, which is why the name is spelled out so plainly. The localization files were already in place: the German bundle renders it as “o, dein Assistent rund um die Uhr,” with Japanese, Hindi, and Korean bundles keeping the lowercase o untranslated. One X user counted the string across 63 languages — preparation that looks far more like a global release than a one-market test.

Then the configuration. On 26 September, TestingCatalog — which tracks unreleased features — posted a screenshot of ChatGPT’s dynamic configuration showing a value block containing "email_suffix": "-o" and "display_name": "o". The suffix is the only capability hint in the config itself: it suggests o will operate with its own identity, likely handling email as a separate, named sender rather than acting invisibly on the user’s behalf. That reading is reasonable but not documented; how the suffix is applied remains unknown.

Finally, the humans. Late on 25 September, Codex and ChatGPT Work suffered an outage. Thibault Sottiaux, described by Wired as OpenAI’s head of core products, posted “o no :(” at 23:08 UTC, then followed an hour later 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. It may be nothing more than a joke — but it is the weakest evidence in the set, and prediction-market traders treated the whole package seriously anyway.

What an always-on assistant actually is

The name is the biggest clue about the product’s shape. OpenAI’s existing agent features in ChatGPT are modes and tools you pick for a task, and they 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” describes availability, not a single job.

The technical foundation appears to already exist in OpenAI’s open-source Codex repository, in a lineage of commits the company has been merging since August:

• 26 August (#40799): a “Persistent” reasoning-effort mode that keeps working instead of stopping at a time limit.
• 26 August (#40942): a current-time reminder and an interruptible clock.sleep tool for persistent turns, so an agent knows what time it is and can wait without ending the task.
• 27 August (#41050): proactivity instructions — “look for useful follow-ups that directly support the completed work,” prioritized using past instructions, “not to infer new authorization.”
• 28 August (#41243): a sleep_tool setting with two modes, model_driven and always_on.
• 25 September (#48100): an HTTP client for remote agent message boards — shared noticeboards where several agents can post findings and receive live notifications, with previews capped at 150 characters.

The proactivity instructions are the most telling. The same file sets hard limits: “Persistence does not broaden that scope,” and any follow-up that would “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 check-in rhythm: “often 1–3 minutes for active near-term work,” backing off when progress is slower. Wired once asked OpenAI about Persistent mode and was told it was being tested with no immediate launch plans; a month later, a named always-on assistant sits in the pricing strings.

Pricing: a $100 floor for something that never sleeps

The bundle describes o as a “Benefit shared by all Pro tiers.” The tier structure visible in the code:

Plan	Monthly price	Code label	Always-on o
Plus	$20	not a Pro tier	not listed
Pro Lite	$100 (screenshot)	“Standard”	included
Pro	$200 (Help Center)	“More usage”	included
Pro Max	$500 (reported)	“Max usage”	included

Compute is the obvious reason for the $100 floor. An assistant that checks back every few minutes for days burns far more tokens than one that answers and stops — and OpenAI has been capacity-constrained enough this month to pause new ChatGPT Pro subscriptions on the back of Astra demand.

The competition is not waiting

OpenAI is not first here. Meta launched Muse on 8 September as a personal agent that can make suggestions unprompted, running on a secure virtual machine and handling travel and shopping through WhatsApp. xAI shipped Grok Bot on 11 August as a desktop app with named AI teammates doing background work. Anthropic’s Conway, an always-on agent for Claude on iOS, was reported in development in July. TestingCatalog’s own verdict: “Meta set a super high standard with its Muse assistant. OpenAI will need to cross a very high bar.”

The timing is not subtle. DevDay takes place Tuesday, 29 September at Fort Mason in San Francisco, with Sam Altman giving the opening keynote. TestingCatalog expects o to be announced there, alongside the Managed Agents and agent-environment developer products it previously reported. Commentators also expect a “fast mode” served from Cerebras hardware tied to the $500 plan, and the shared agent message boards that landed in Codex on Friday. According to Octagon, which tracks prediction markets, Kalshi’s contract on OpenAI releasing a personal AI agent before 1 January 2027 jumped from 24% to 78% after the leak.

The caveat, and what to watch

Nothing OpenAI has published describes o’s features; everything above is assembled from leaked strings, public Codex commits, and third-party code readings. The four open questions the leak does not settle: where o runs (a cloud computer, per one commentator’s claim), how exactly it handles email, what it may do without explicit approval, and which countries get it first.

For a lab whose agents have spent September making headlines for browsing US government websites without authorization, an assistant that never clocks off is a bold product to ship. The authorization-scoping language in the Codex commits — persistence never broadens scope — will be the line to watch when the keynote lands. If o ships Tuesday with those guardrails intact, it is a new category of consumer software. If it ships without them, it is a very busy week for everyone else.