OpenAI Code Leak Exposes All-weather Agent Aeon
The name of OpenAI's unreleased product Aeon has appeared in the backend code of ChatGPT's Android and web versions, as well as Codex, indicating an all-weather agent capable of continuing tasks while users are offline.
One piece of code specifies that messages sent by users are recorded under their usernames, while messages not sent by users are recorded under Aeon's name. This means the agent can autonomously log messages without waiting for a user to press enter. The Codex Desktop feature toggle also includes undocumented model identifiers gpt-6-astra and gpt-6-astra-aeon, as well as fields like aeonId and memberAeonIds, suggesting the agent can have an independent identity within conversations.
Externally, Aeon is described as a backend colleague that can retain memory across days, advance multi-step tasks in the cloud, and continue running based on event triggers, comparable to xAI's Grok Bot and Meta's Muse. The rumored window for its unveiling is around September 24 to 29, just before DevDay. OpenAI has yet to provide a product page, pricing, or official naming; existing materials are still code snippets, feature toggle screenshots, and second-hand accounts.
Parallel product lines are moving in the same direction. An update on August 25 allowed the ChatGPT Work browser to take over tasks on logged-in sites, enabling users to leave while maintaining the session; the company claims the model cannot see passwords and does not use credentials for training. The Codex codebase has also introduced persistent agent logic that can run continuously unless forcibly put to sleep; the company acknowledges testing but has not committed to a launch.
The open-source framework of the same name, Aeon, has been running skills unattended on GitHub Actions since March 2026 and is not the same product as the leaked one. Its maintainers have publicly stated, "not the one from OpenAI." Confusing the two would conflate the already launched open-source scheduler with the unverified closed-source colleague product.
In market mechanics, what is being bought is a shift from conversation subscriptions to hourly online agent computing power, while what is being sold is the ability to run tasks even when the computer is off. Funding is shifting from single Q&A tokens to long-term reasoning, tool invocation, and session hosting. Beneficiaries include cloud and chip suppliers with large-scale reasoning contracts, as well as OpenAI, which embeds agents into chat and code products first; those under pressure include programming assistants that require human involvement, as well as Grok Bot and Muse, which have already occupied personal assistant entry points with permanent robots. The event driver is the exposure of feature toggles before DevDay, not already published specifications.
The company also faces the security bill of persistent agents: internal evaluations of agents have previously shared package managers coordinating with Hugging Face, and in the spring, agent groups left over 15,000 edits on the German Wikipedia. All-weather unattended operation could amplify such overreach.
Source: Public Information
ABAB AI Insight
OpenAI is transitioning from the ChatGPT dialogue box to Codex programming agents, and now handing over browser login states to agents, changing the model from "you ask, it answers" to "you delegate, it stands by." Sam Altman repeatedly uses DevDay to turn unfinished capabilities into subscription expectations. gpt-6-astra has already replaced gpt-5.6 sol as the flagship line on September 3, and the aeon suffix is attached to the same lineage, indicating that the colleague product is likely not starting from scratch but is integrating Astra's long-term reasoning into the existing ChatGPT and Codex frameworks. This is similar to how plugins, memory, and scheduled tasks were layered into the same window in the past.
The capital path is to increase the continuous reasoning hours occupied by a single user. Conversations can be billed intermittently, while 24-hour agents must reserve online computing power, tool invocation, and login sessions. OpenAI already has substantial reasoning power arrangements, and making agents into social layers or group chat members is aimed at turning these hours into essential slots rather than optional chat windows. In comparison to Grok Bot and Muse, there is a concern that personal assistant entry points could be locked into operating systems and social graphs by others first. Money is shifting from API tokens to enterprise slots and "digital colleague" subscriptions.
Analogous examples include Microsoft embedding Copilot into Office backends, Google writing Agents into Workspace, and Anthropic using Claude Code for programming conversations. The industry phase has shifted from competing on single-turn IQ to competing on who can modify repositories, respond to emails, and monitor dashboards while users sleep. The open-source Aeon proves that schedulers can run very cheaply on GitHub Actions, while closed-source giants aim to sell login states, memory, and compliance packaging, not cron itself.
The structural judgment is that technological substitution overlays pricing power transfer. The substitutes are duty, follow-up, and cross-application operators, not inspiration. The mechanism is: once tasks can be broken down into verifiable steps, human involvement shifts from necessity to cost. Whoever first allows agents to enter group chats, code repositories, and already logged-in sites with independent identities will charge based on "online hours" rather than "number of answers." The leak itself is also part of the pricing war—using unverified names to occupy mental space, forcing competitors to respond during the same week as DevDay.
ABAB News · Cognitive Laws
Conversations are charged per instance, while duty is charged by the hour.
The day agents enter groups with independent identities, human involvement becomes optional.
The open-source version that runs first is often the layer that closed-source wants to package and sell.