OpenAI Code Strings: Surprising Clues to Smart Dots Agents

OpenAI code strings found in recent ChatGPT builds suggest that the company’s upcoming personal AI agents will be called Dots. The key line, “Dots, your remarkably capable agents,” surfaced at 04:56 UTC on 29 September 2026, about twelve hours before OpenAI’s DevDay keynote in San Francisco. Further strings describe agents you can text, call, message on Slack or email, which run on their own computer and can buy things with your approval.

We covered what those strings say the product will do in our earlier report on OpenAI Dots. This article looks at the evidence itself. Code strings are now the main way the public learns about AI products before launch, and they are easy to over-read.

Below, we explain what code strings are and how leakers find them, what each Dots string proves and what it does not, how the product’s name moved from “aeon” to “o” to “dots” in eight days, and how often string leaks this month turned into real launches.

Table of contents

What the OpenAI Code Strings Say About Dots
What Code Strings Are and Where Leakers Find Them
Reading the OpenAI Code Strings One by One
The Name Trail in the OpenAI Code Strings: Aeon, o and Dots
How Reliable Are OpenAI Code Strings?
What OpenAI Code Strings Cannot Tell You
How Businesses Should Treat OpenAI Code Strings
When the OpenAI Code Strings Will Be Tested
OpenAI Code Strings FAQs
References

What the OpenAI Code Strings Say About Dots

The first post came from the leaker M1 (@M1Astra). It quoted the product line and summarised the rest: “You can text it, call it, Slack it, and email it, and it can buy things for you with your approval. Each dot gets its own 3D character you design.” Eleven minutes later M1 added a second line: “It designs itself from what it knows about you.”

The longer set of strings

At 05:07 UTC, ChrisGPT (@ChrisGPT) posted more of the OpenAI code strings, gathered “in collaboration with” M1. They included “Your dot knows when to take action, and when to ask for approval. Add custom rules for more control,” “Your research is ready,” and “Your dot’s computer.” A follow-up post read simply: “Only on Pro.” TestingCatalog relayed the news at the same minute, saying Dots “will be the name of upcoming AI agents from OpenAI, according to new strings that have surfaced in recent builds.”

A rename still in progress

One small detail matters more than it looks. Asked about the earlier name, M1 replied that “‘Delete o?’ shows up in one spot. Everywhere else it’s dots.” In other words, the OpenAI code strings still contained one leftover prompt from the product’s previous name. That is typical of a rename made late in development, and it is strong evidence that “dots” replaced “o” rather than the other way round.

What Code Strings Are and Where Leakers Find Them

A code string is a piece of text inside an app that is meant to be shown to users: a button label, an error message, a setting, a marketing line. Web apps like ChatGPT ship their interface as JavaScript files that every browser downloads, and those files often contain text for features that are built but switched off. Mobile apps carry the same kind of text inside their install packages.

Interface copy, configuration and assets

Not every piece of evidence behind the OpenAI code strings is equal. Leakers such as M1, ChrisGPT, Tibor Blaho and TestingCatalog work with several kinds of material, and each deserves a different level of trust.

Type of evidence	Example from the Dots leak	What it usually proves	Weight
Internal codename	“Aeon” fields in Codex code	A project exists; says little about the launch name	Low for naming, medium for existence
Configuration value	“O” as a display name and “-o” as an email suffix	How the feature is wired, at that moment	Medium
User-facing copy	“Dots, your remarkably capable agents.”	What OpenAI planned to tell users	High, until it is edited
Plan or upgrade page text	“o” listed as a benefit of the $100 Pro plan	Intended packaging and pricing	High, but plans change
Design assets	Sheets of 3D character shapes and eyes	That a visual feature was designed	Medium
Announcement strings	The redacted “Spaces” notice found the same night	Launch copy exists	High for timing, low for detail if redacted

Why companies ship text before launch

Products are usually built behind feature flags, switches that keep a finished feature hidden until launch day. The interface text is written, translated and bundled weeks in advance, so it can sit in public files long before the switch flips. That is why OpenAI code strings so often appear days before an event, and why the same strings can quietly change if marketing picks a new name.

How the OpenAI code strings for Dots surfaced

The OpenAI code strings for Dots moved quickly from one account to the next. M1’s post appeared at 04:56 UTC, ChrisGPT’s longer set and TestingCatalog’s relay both at 05:07, and TestingCatalog’s analysis linking Dots to “aeons” at 05:22. ChrisGPT then added: “me and @M1Astra are very excited for tomorrow and in no means want to steal the show so we will leave the rest to OpenAI!” That suggests some strings were held back.

Reading the OpenAI Code Strings One by One

The table takes the main OpenAI code strings in turn and separates what each one shows from what it leaves open. The gap between the two columns is where most leak coverage goes wrong.

String	What it shows	What it does not show
“Dots, your remarkably capable agents.”	A consumer-facing name and tagline were written	That the name will survive launch
“You can text it, call it, Slack it, and email it”	Four planned contact channels	Which channels work on day one, or in which countries
“It can buy things for you with your approval”	Purchases are designed to need approval	Spending limits, payment methods or liability
“Your dot knows when to take action, and when to ask for approval. Add custom rules for more control.”	The agent decides when to ask, and users can add rules	What the default rules are, or who can override them
“Your dot’s computer”	Each dot has its own computer, likely a cloud machine	Its specification, or how long it keeps running
“ChatGPT can securely store your passwords to log into websites when using its browser.”	A password store for ChatGPT’s browser is planned	That it is limited to Dots, or how it is protected
“Only on Pro”	A Pro-only launch was planned when the text was written	Which Pro tier, or whether it stays Pro-only
“Delete o?”	The product was recently called “o”	Whether “o” survives as a separate product

The approval string deserves the closest reading

The line “Your dot knows when to take action, and when to ask for approval” is the most consequential of the OpenAI code strings, because it puts the model in charge of deciding when to ask. That design choice will matter more than any name. It arrives the day after the Wall Street Journal reported that OpenAI had cancelled its GPT-6.1 Astra release over behaviour that included acting without permission.

The password string may be a separate feature

The password line mentions ChatGPT and its browser, not dots. It may describe a general ChatGPT feature that Dots uses, rather than something unique to the agents. Reading all the OpenAI code strings in a leak as parts of one product is a common mistake.

The Name Trail in the OpenAI Code Strings: Aeon, o and Dots

The most useful thing the OpenAI code strings reveal is how fast the product’s name changed. Over eight days, public reporting tracked three names for what appears to be the same always-on agent.

Date (2026)	Name	Where it appeared
6 to 22 April	“Hermes”, shipped as workspace agents	TestingCatalog reports on an agent builder, then OpenAI’s launch
21 September	“Aeon”	Posted by Tibor Blaho as The Information reported OpenAI’s answer to Grok Bot
22 September	“Aeon” fields	Agent name and appearance fields found in Codex code by ChrisGPT
26 September	“o”	Display name, “-o” email suffix and a $100 Pro plan benefit, reported by TestingCatalog
28 September, 01:32 UTC	“. . . . . .”	A post by OpenAI’s Thibault Sottiaux consisting only of dots
29 September, 04:56 UTC	“Dots”	User-facing copy found by M1, with one leftover “Delete o?”

Aeon in the OpenAI code strings

“Aeon” first appeared publicly on 21 September, when Tibor Blaho posted the single word alongside a report from The Information that OpenAI was developing an answer to Grok Bot. The next day, ChrisGPT reported “Aeon” fields in Codex code covering an agent’s name and appearance. TestingCatalog later connected Aeon to the custom agents OpenAI launched for business accounts on 22 April.

“o” on the upgrade page

On 26 September, TestingCatalog reported “references to ‘O’ across ChatGPT’s configuration, including ‘O’ as a display name and an email suffix configured as ‘-o’.” It added that the agent “was also recently referenced as a benefit on the upgrade page for the $100 Pro plan.” That upgrade-page text is the strongest evidence in the whole trail, because it was briefly shown to real users rather than hidden in code.

The six-dot teaser

In hindsight, OpenAI may have hinted at the change itself. At 01:32 UTC on 28 September, Thibault Sottiaux posted a message consisting of nothing but six dots. After the Dots strings surfaced, one reply to TestingCatalog linked back to it with the comment “now it all make sense.” By about 13:00 UTC on 29 September, the post had drawn more than 2.5 million views, far more than any of the leaks. The chart compares them.

Views on X at about 13:00 UTC, 29 September (bar length relative to the six-dot post)

Sottiaux’s six-dot post 2,548,306

M1’s Dots strings 222,478 (8.7%)

ChrisGPT’s longer set 168,986 (6.6%)

TestingCatalog’s relay 64,356 (2.5%)

How Reliable Are OpenAI Code Strings?

The honest answer is: reliable about existence, much less reliable about names and details. September 2026 gives an unusually clean record, because several leaks resolved within days. The table covers the ones that bear on AI product launches, including two from Anthropic for comparison.

Leak	First reported	Outcome	Time to outcome
“Hermes” agent builder	6 April	Shipped 22 April as workspace agents	16 days, under a new name
GPT-6 Sol and Luna	22 September, 12:07 UTC	Launched the same day	About 9 hours
Claude Opus 5.5 (Anthropic)	22 September, 02:41 UTC	API rollout the same day	About 14 hours
Claude Sonnet 5.5 (Anthropic)	27 September, 21:48 UTC	Rolling out the next day	About 20 hours
“o” always-on agent	26 September	Strings renamed to dots	3 days, under a new name
ChatGPT Pro Max at $500 a month	24 September	Not launched by 29 September	Pending
Free, Prototype and Accelerate Platform plans	23 September	Not launched by 29 September	Pending, a DevDay candidate

Model leaks resolve fastest

The three model leaks all resolved within a day. The chart shows the gap between each leak report and the launch report, measured in hours from the timestamps above.

Hours from leak report to launch report, September 2026 (bar length relative to 24 hours)

GPT-6 Sol and Luna 9.0 hours

Claude Opus 5.5 13.7 hours

Sonnet 5.5 (Anthropic) 20.2 hours

Product names are the weakest part

Two of the seven leaks above changed name before launch: “Hermes” became workspace agents, and “o” became dots within three days. Model names were stable, because they are fixed well before launch. So when OpenAI code strings name a product, treat the name as provisional. Product names are marketing decisions, and they can change until the last week.

Plans can change entirely

Leaks describe intentions, and intentions change. GPT-6.1 Astra is the clearest recent example. The Wall Street Journal reported on 28 September that OpenAI had planned to launch it “in the coming days or weeks, aiming for an October debut,” then cancelled it after safety testing. Any OpenAI code strings found for a product like that would have pointed to a launch that never came.

What OpenAI Code Strings Cannot Tell You

Strings describe the interface. They say little about the terms, the model or the limits, which are exactly what a business needs before approving a tool.

Price and plan

“Only on Pro” and the $100 Pro plan reference suggest a Pro launch, but OpenAI now sells more than one Pro tier and has been preparing a higher $500 Pro Max plan. The OpenAI code strings do not say which tier gets Dots, or whether running several agents draws on a usage allowance.

Model, availability and safety defaults

Nothing in the OpenAI code strings names the model behind Dots, the countries where it launches, or the default approval rules. Those details decide whether a personal agent is safe to let near work accounts, and they will only come from OpenAI’s announcement and terms.

Whether businesses get it at all

The strings describe a consumer product. OpenAI’s business customers already have workspace agents, with admin controls and approvals. Whether Dots comes to Business and Enterprise plans, and with what controls, is not in the OpenAI code strings at all.

How Businesses Should Treat OpenAI Code Strings

Leaks are useful early warnings, but they are a poor basis for decisions. A simple approach keeps them in their place.

Plan, do not buy

Use leaks such as the OpenAI code strings to start policy conversations, not procurement. If a leak points to a tool your staff will want, such as a personal agent that can read email, begin thinking about the rules now. Do not commit budget or change contracts until the launch terms are published.

Update policies before personal agents arrive

The Dots strings describe agents that can join Slack, send email and log into websites with stored passwords. Company credentials stored in a personal agent sit outside your identity controls, so this is a cybersecurity question as much as a productivity one. Our guide to AI employees and autonomous agents covers how to set rules for agents that act on staff’s behalf.

Keep a short watch list

Track the handful of leaks that would affect you, and note what would need to be true for each to matter. OpenAI code strings will keep appearing before every event, so a written watch list saves re-reading each leak from scratch. For Dots, that list might read: business availability, admin controls, approval defaults, data retention and price.

Verify on launch day

When the announcement comes, compare it with the leak. Anything the strings suggested but the launch omits should be treated as not shipped, however confident the pre-launch coverage was. Our AI strategy work helps teams turn launch-day news into a considered decision.

When the OpenAI Code Strings Will Be Tested

The test comes at DevDay. OpenAI’s schedule puts the opening keynote, “featuring Sam Altman,” at 10:00 a.m. Pacific Time on 29 September, which is 17:00 UTC and 18:00 in London. OpenAI had already teased “20+” new launches the evening before, according to TestingCatalog.

What would confirm the strings

A product called Dots, reachable by text, phone, Slack and email, with its own computer and approval rules, would confirm the core of the leak. A different name with the same features would confirm the product but not the branding, which is the pattern “Hermes” followed in April.

What would contradict them

If OpenAI presents “o” as the assistant and dots as agents it manages, both names could survive. If neither appears, the OpenAI code strings may describe a product held back for later, as has happened to other features found in OpenAI’s code.

OpenAI Code Strings FAQs

What are OpenAI code strings?

They are pieces of user-facing text, such as labels, messages and marketing lines, found in the files that ChatGPT’s apps download. They often describe features that are built but not yet switched on.

What do the OpenAI code strings say about Dots?

They describe “Dots, your remarkably capable agents”: personal AI agents you can text, call, Slack or email, that run on their own computer, ask for approval before buying things, and each have a 3D character.

Has OpenAI announced Dots?

Not before the DevDay keynote on 29 September 2026. The information comes from strings posted by M1 and ChrisGPT and relayed by TestingCatalog. DevDay may confirm, rename or delay it.

Are string leaks usually accurate?

They are usually accurate about what exists, and the September model leaks all resolved within a day. They are less reliable about names, prices and timing, which can change until launch.