FULL STORY

OpenAI DevDay 2026: From Leaks to Launch

Ahead of OpenAI DevDay, leaks revealed the persistent agent "o" and Dots. At the September 29 event, Codex went fully cloud-native and dots launched amid a botched demo and mixed reception.

2026-09-25 ~ 2026-10-03 · 14 episodes · 156 posts

Episode 1 · Rumors swirl ahead of OpenAI DevDay: always-on agents and new models (2026-09-25, 4 posts)

Ahead of OpenAI's September 29 DevDay, community speculation centers on an always-on agent alongside Sol 6/Luna 6, plus hints like astra minor, ultrafast mode, a possible Pro Max tier, and voice updates, with mixed expectations on whether stronger models will appear.

Episode 2 · OpenAI's always-on agent "o" seemingly leaked by its own staff (2026-09-26, 13 posts)

OpenAI appears to have accidentally leaked the product name of its first always-on agent: a lowercase "o". Between Sept 26–28, multiple independent sources found references to "o" in ChatGPT web code, internal configs, and some users' upgrade pages; OpenAI engineer Thibault Sottiaux's remarks on X suggested an employee slip. The community considers the clues credible, with the product expected to debut at DevDay on Sept 29. Blogger VraserX predicts it will be OpenAI's biggest launch of the year: an always-on, fully autonomous agent that taps user context via memory and plugins and keeps working, rather than a chatbot opened on demand.

Confirmed

  • The product name "o" (lowercase) genuinely appears in ChatGPT internal configs and web code, independently reported by @testingcatalog, @markk, @CodexResets1 and others; per 新智元, developers dug out explicit comments like "Lowercase o is the assistant's product name" from ChatGPT core code, treated as a "code smoking gun."
  • "o" appears as a display name in internal config with an "-o" email suffix, suggesting the assistant will support email handling with its own mailbox identity.
  • Per Thibault Sottiaux's X remarks, the codename was seemingly leaked accidentally by OpenAI staff; the community largely believes the clue is real and has been joking about it.
  • On Sept 28, observer kimmonismus noted Tibo (Tibo Louis-Lucas) had hinted at the internal "o" project for two days, calling it significant for OpenAI and predicting an imminent Astra update.

Unconfirmed

  • Reddit user @1411337 claims the always-on assistant is built on a Google Astra variant called "Aeon," optimized for long-running tasks; @CodexResets1 likewise says DevDay will reveal a long-running Agent codenamed "Aeon," similar to Grok Bot or Manus, working for hours, days, or weeks, possibly built on Astra. Both are unverified single-source claims; the relationship between "o" and "Aeon" remains unclear.
  • @ExpertAnnual19 says "o" will be open to all Pro tiers and include Fast Mode — unconfirmed officially.
  • Timing: @markk says Tuesday's DevDay; @testingcatalog reports a possible Sept 29 DevDay reveal in Sam Altman's opening keynote; @CodexResets1 says Thursday. @xiaohu and 新智元 echo the DevDay expectation; specifics remain rumor-level.
  • Positioning: multiple posts describe it as a Grok-like always-on bot (Muse also mentioned), 24/7 online and cloud-dependent; @VraserX further predicts a fully autonomous agent accessing user context via memory and plugins, calling it OpenAI's biggest release of the year — all blogger speculation, not official.
  • @Dullydude dug up a Sept 20, 2024 report that OpenAI might rebrand to a single letter "O," speculating a link to the new agent's name — pure conjecture.

Why it matters

  • If true, this would be OpenAI's first consumer-facing always-on agent, a shift from conversational assistant to a continuously online agent that proactively handles tasks like email, directly rivaling xAI's resident Grok bot.
  • The codename leaked by its own staff, appearing repeatedly in product code with explicit comments, suggests the product is in late-stage configuration ahead of launch — Sept 29 DevDay is worth close attention.

Episode 3 · OpenAI DevDay Enters Final Countdown, Livestream Confirmed (2026-09-27, 8 posts)

OpenAI DevDay is in its final countdown. On Sept 27 the official OpenAI dev account posted a 72-hour teaser saying the team has been building and "it's time to show what we've made," accompanied by a preview video. On Sept 29, OpenAI staff member reachvb confirmed the event begins in 24 hours and will be livestreamed. The company also showcased Codex Ambassadors visiting its headquarters. Specific announcements remain undisclosed, making DevDay one of the most closely watched near-term events in the developer community.

Confirmed

  • The official OpenAI dev account posted the 72-hour countdown with a preview video teasing upcoming developer updates.
  • OpenAI's reachvb confirmed a 24-hour countdown and a livestream for the event.
  • Codex Ambassadors were shown visiting OpenAI HQ; developer tools and model updates are expected.
  • Developer-community account danielmac8 reshared the teaser, calling it "The day of the Dev."

Unconfirmed

  • AI account agentnative claimed this DevDay could be OpenAI's biggest single-day launch in a long while and urged developers to clear their schedules — hype, not officially verified.
  • altryne noted earlier leaks about a new personal-assistant product codenamed Aeon (or O), positioned against Character.ai; unconfirmed by OpenAI.

Why it matters

  • DevDay is OpenAI's flagship release window for developers; the sustained official teasers and confirmed livestream typically signal a concentrated wave of tooling and platform updates worth close attention.

Episode 4 · Investor Predicts OpenAI Dev Day: 50% Price Cut and Bigger Models (2026-09-27, 2 posts)

Investor Bindu Reddy predicts OpenAI's Dev Day will bring Astra 6.1/6.5, a promise of the larger next-gen model Bel, and a 50% price cut for Astra or Sol 5.6, while Peter Yang claims the rumored Sol 6 is actually Anthropic's Opus 5.

Episode 5 · OpenAI DevDay agenda leaks across multiple channels, Codex platform plays take center stage (2026-09-28, 7 posts)

OpenAI DevDay, scheduled for next Tuesday, is widely seen as one of the most important AI events of the year. Ahead of the event, its agenda has leaked through multiple channels: Reddit user @PsychologicalJob614 posted schedule screenshots, @TheOyinbooke's screenshots suggest more sessions may be added the next day, and @sloppenheimer found the agenda quietly appearing inside the ChatGPT app, indicating it is finalized and an official announcement may be imminent. The official agenda only promises a keynote unveiling "what's next" plus 22 technical breakout sessions.

Confirmed

  • The official agenda includes one keynote and 22 technical breakout sessions, with no specific product list disclosed.
  • Schedule screenshots have circulated (@PsychologicalJob614, @TheOyinbooke), and the agenda appearing in the ChatGPT app (@sloppenheimer) corroborates the leaks.

Unconfirmed

  • @testingcatalog reports OpenAI will build platform capabilities into Codex, letting users develop plugins, agents, and apps from a dedicated DevDay page.
  • Rumors relayed by @VraserX include always-on agents with persistent memory, a first demonstration of continual learning, autonomous AI research interns, and a possible first public showing of io hardware.
  • @johnseach extracts from session titles a hint of long-running coding agents: a single Codex run lasting 1000+ hours.

Why it matters

If these rumors hold, Codex would graduate from a coding tool to a development platform, while persistent memory, continual learning, and very long-running agents would mark a major leap in OpenAI's autonomy push; io hardware could reveal OpenAI's hardware ambitions for the first time. The agenda surfacing in the ChatGPT app suggests preparations are nearly complete—treat the keynote's official announcements as authoritative.

Episode 6 · OpenAI DevDay 2026 Kicks Off With 20+ Launches and a Mystery Reveal (2026-09-28, 44 posts)

OpenAI DevDay 2026 took place on Sept 29 at Fort Mason, San Francisco, billed as the largest DevDay ever. The keynote at 10 AM PT, led by Sam Altman, Romain Huet, Tejal Patwardhan and Holly Li, promised 20+ launches in a single day plus first demos of the latest breakthroughs. Altman personally teased "We have found a new thing," with all details held for the keynote.

Confirmed

  • The official DevDay trailer promised "over 20 new releases," noted by testingcatalog, Polymarket, and others; the developer account shared the full agenda with a 10 AM keynote, noting in-person registration was closed but the livestream was open to all
  • OpenAI posted a cryptic "Get ready" tweet; with Anthropic's recent Opus 5.5 and Sonnet 5.5 releases as competitive pressure, observers widely expected a major model or product launch
  • Altman teased "We have found a new thing" and, on event day, said the livestream would begin in two hours with "some good stuff" prepared; the vague phrasing triggered heavy community speculation
  • The livestream page confirmed Altman, Romain Huet, Tejal Patwardhan and Holly Li presenting 20+ new features and first-ever demos; testingcatalog ran a live thread and applesjimmy relayed updates from friends on site
  • OpenAI employee Platon Mazarakis previewed a year's work on rethinking "how work happens in the agentic era" and solicited Chat/Codex improvement ideas; jxnlco signed a speaker release form, both hinting at new product launches
  • A Polymarket contract on OpenAI shipping consumer hardware this year saw over $440K in volume with a 33% implied probability

Unconfirmed

  • A third-party leak relayed by markk claimed three launches: a persistent Agent "OpenAI o", Astra 6.1, and a unified ChatGPT-Work integration
  • testingcatalog reported a new product called "Dots" and flagged codename "O"; other rumors mentioned model codenames like "bel" and "astra minor" — all unverified

Why it matters

  • The 20+ launch scale suggests a rare same-day push across models, agents, and the developer ecosystem
  • Altman's personal teaser, employee hints about agentic work reimagined for ChatGPT/Codex, and leaks about a persistent agent would mark OpenAI's shift from chat product to an always-on workflow platform if confirmed
  • Prediction markets pricing 33% odds on consumer hardware reflect expectations beyond pure software

24 more related posts →

Episode 7 · Leaked OpenAI DevDay Rundown Lists 20+ Launches: Always-on Agents, Codex Cloud (2026-09-28, 8 posts)

Days before DevDay, multiple accounts circulated a purported OpenAI launch list covering 20+ products and features, with the overarching narrative that ChatGPT evolves from a chat box into a native platform for "humans + agents." None of the list items are officially confirmed; only the DevDay event itself is (@CodexResets1 confirmed the timing). @koltregaskes also grouped these rumors alongside possible same-week releases of Anthropic Sonnet 5.5 and Google Gemini 4.

Confirmed

  • OpenAI DevDay will take place soon (@CodexResets1 confirmed the event timing).

Unconfirmed

  • Models and subscriptions: GPT-6.1 Sol (@eyishazyer says near-Astra intelligence at roughly one-fifth the price), GPT-6.5 Astra, GPT-6 Cyber, GPT-6 Aeon, ChatGPT Pro Max (compiled by @iamfakhrealam).
  • Always-on 24/7 agents, aka "Dots" or the "O" agent, mentioned across @kimmonismus, @eyishazyer, @CodexResets1, @koltregaskes and @VraserX, with @VraserX describing Dot as a persistent autonomous agent.
  • Spaces, a persistent shared AI workspace, per @VraserX and @kimmonismus.
  • Developer-side items: Codex in the cloud, persistent Codex agents, multiple Codex versions, Agents API, OpenAI marketplace, and OpenAI's in-house "jev".
  • Ultrafast inference mode, which @kimmonismus suggests may be limited to the $500 tier; @koltregaskes and @CodexResets1 cite a first $500-600 ChatGPT subscription plan, with @CodexResets1 also flagging a hardware debut.
  • @NFTChen relays a claim that OpenAI accidentally leaked its own DevDay 2026 recap document (20+ updates)—itself unverified.

Why it matters

If the list holds, it marks OpenAI's strategic shift from a conversational product to an agent platform, and the premium tier plus Ultrafast inference could reshape ChatGPT's product form and pricing. The same items (always-on agents, Codex cloud, $500 tier, Ultrafast, Spaces) appearing across independent sources raises credibility, but caution is warranted before any official announcement.

Episode 8 · Ahead of OpenAI DevDay, Wrapper Startups Brace for "D-Day" (2026-09-28, 7 posts)

Ahead of OpenAI DevDay, the AI community flooded X with jokes and speculation reflecting deep anxiety in the downstream ecosystem about being "absorbed" or "replaced" by the platform. No official announcements had been made yet, but the pre-event sentiment itself became the story.

Confirmed

  • e/acc founder beffjezos jokingly called DevDay "D-Day" for AI wrapper startups, implying OpenAI's releases tend to wipe out products built on its API; the meme spread widely, with users like MickeySteamboat resharing and himanshustwts summarizing it as "one API change and a batch of startups gets cooked"
  • David Khourshid (DavidKPiano) and chip executive idanbeck both predicted only two outcomes: a model smarter and cheaper than Opus 5.5, or "mild disappointment" — reflecting high expectations; a reply to idanbeck joked his only two outcomes were eating lunch or forgetting to
  • Blogger iScienceLuvr half-jokingly urged users to burn through Codex quota before DevDay, predicting an "inevitable" reset, revealing expectations that subscription quotas refresh with the event
  • Josh Bickett summarized that startups are waiting to see whether they'll be accelerated or replaced, while Codex subscribers weigh whether to cancel

Why it matters

These jokes expose the structural tension between model vendors' launch events and the downstream app ecosystem: each major platform update can swallow the生存空间 of a cohort of wrapper startups and reshape subscribers' payment decisions, making DevDay a bellwether for the AI application layer.

Episode 9 · Report: OpenAI to Launch Agent Product Dots to Rival Grok and Meta (2026-09-29, 3 posts)

Leakers including ChrisGPT and @M1Astra claim OpenAI is set to launch a personal agent product codenamed "Dots," rivaling Meta Muse and Grok Bot, featuring independent VMs with password storage, 3D personas, multi-channel outreach, and shopping capabilities. The details remain unverified.

Episode 10 · OpenAI DevDay: Codex Goes All-In on Cloud as Live Demos Falter (2026-09-30, 41 posts)

At its Sept 29 DevDay, OpenAI announced that its coding agent Codex is moving fully to the cloud: Codex Cloud offers configurable, reusable cloud environments with repos, dependencies, scripts and settings pre-provisioned, letting agents keep working after users close their laptops, with handoff to phones, and is rolling out to Plus, Pro, Business and Enterprise plans. An Agents API with computer use entered preview beta. Over 20 products were announced, with the through-line of AI shifting from answering questions to doing work—but consecutive live demo failures and disputed replay edits became the event's other headline.

Confirmed

  • Codex cloud environments launched: reusable, pre-provisioned environments that reduce setup time and keep agents running after laptops close
  • Cloud environments available to Plus, Pro, Business and Enterprise subscribers, selectable when starting any task across Codex entry points
  • Official demo shows tasks started on a laptop, resumed on a phone; environments include repos, network config and secrets, and can be shared with teams
  • Lenny Rachitsky reported Claire Vo and Codex lead Thibault Sottiaux performed a 'Live global Codex reset' on stage
  • Agents API entered preview with computer use, beta now open on the API side
  • Thibault Sottiaux said that once you set up a cloud environment, you can never go back to local laptop development
  • Other DevDay releases include always-on personal agent Dots (Astra-powered, usage-based billing) and ChatGPT collaboration spaces; Sam Witteveen and David Ondrej published recap videos; jonkragh flagged three underappreciated signals: plugin distribution window, Codex security and the Decisions API
  • On-stage incidents: Dots failed to respond in its first live demo (host joked 'Dot is not feeling well this morning'); a Codex voice chat demo also failed; Dots demos failed twice

Unconfirmed

  • ns123abc claims the official livestream cut audio at the exact moment of demo failure and posted a phone recording as evidence; Community Notes contributors dispute this, and OpenAI has not commented
  • Users report the failed segments are missing from the official replay, but deliberate editing is unverified
  • Posts also claim searching Dots redirected to Grok Bot and CFO Sarah Friar misspoke Muse as Dots on TV—unverified side details

Why it matters

  • Cloud agents free coding tasks from local hardware and network constraints; venturetwins quipped that OpenAI productized the trick of leaving a laptop running in a car on a power bank, while cneuralnetwork called Codex 'completely cloud native', suggesting the pure-local era may end
  • Phone handoff and team sharing extend cloud dev workflows to mobile and collaboration scenarios
  • Agents API plus computer use lets agents directly operate computer interfaces, expanding automation's reach
  • The demo failures and replay controversy cast a shadow over an agent-centric event: kimmonismus argued the keynote was overstuffed and belonged in workshops; Scoble commented on the stage mishaps; one commenter said the Dot failure prompted them to buy more Meta stock, reflecting market sensitivity to OpenAI's product maturity

21 more related posts →

Episode 11 · OpenAI Breaks Down Its New Agent Stack After DevDay (2026-09-30, 3 posts)

Following DevDay, the Latent Space podcast interviewed OpenAI's Computer Use and API leads about the new agent stack, including superhuman-speed Computer Use, Dots cloud computers and the Decisions API. Turing Post's editor also analyzed which agents will get persistent cloud residency.

Episode 12 · OpenAI Executive Publicly Credits Open-Source Project pidot at DevDay (2026-09-30, 2 posts)

At OpenAI DevDay, executive Thibault Sottiaux publicly credited the open-source project pidot (pid.dev), a moment highlighted by AI scholar Omar Khattab. Observers see it as a signal that OpenAI is seriously investing in a more open ecosystem.

Episode 13 · OpenAI Launches Always-On Agent dots at DevDay 2026 (2026-10-01, 7 posts)

At DevDay 2026, OpenAI launched dots, an always-on agent living inside ChatGPT. It continuously learns what matters to the user, proactively takes work off their to-do list, and runs on its own cloud computer between conversations. Official demos included trip planning and routing user feedback into Slack, quickly sparking developer debate over engineering use cases, mechanics, permission design, and ecosystem impact.

Confirmed

  • Positioning: dots is a persistent agent in ChatGPT with its own cloud compute (@NoSpecific64 notes 400 related configuration details), proactively taking over to-dos.
  • Demos: trip planning and summarizing user feedback into Slack.
  • Mechanics: @Kitchen-Jicama8715 summarized the docs: the dots cloud computer is separate from the user's PC—tasks not needing the local machine keep running after shutdown; local tasks require the device online with the ChatGPT app open. The author also shared a ready-to-use delegation prompt.
  • Engineering access: OpenAI's Rohan Varma said the team is building dots as an engineering orchestration hub, with local access (CUA desktop control, Codex app, built-in browser, local dev environment) and cloud access (relayed by @Dimillian).
  • Anecdote: @lennysan shared that five minutes before the live demo, thsottiaux's agent "Dot" detected a production outage and asked to fix it; thsottiaux quipped that it probably couldn't.

Unconfirmed

  • @lmoroney argues the always-on, cloud-computer part is real, but how permissions, approvals, and accountability are designed when the agent acts on your behalf remains unanswered.
  • The two @lennysan posts disagree on the agent's owner (thsottiaux vs. Simon Willison); details unverified.

Why it matters

  • dots pushes the always-on agent era forward, but @lmoroney contends the hard problems are permissions, approvals, and accountability rather than flashy demos.
  • The pre-demo outage anecdote, though ending in a joke, exposed the gap between agent proactivity and reliability limits.
  • @NoSpecific64, a founder in the same space, candidly assessed the blow to AI agent startups, reflecting platform players squeezing the startup ecosystem.

Episode 14 · OpenAI Unveils Persistent Agent Dots at DevDay, Glitches and All (2026-10-01, 7 posts)

OpenAI unveiled Dots, a persistent personal agent, at its annual DevDay conference attended by over 2,500 people. The launch was seen as the event's biggest highlight, kicking off a personal AI agent war with Meta's Muse (already tested for a month). The live demo suffered multiple glitches that the company attributed to WiFi issues; it later re-recorded and published the demo in a different network environment, showing a user assigning tasks by phone, such as booking compliant, refundable travel and organizing user feedback. Reviewers broadly agree the direction is promising but the product is unfinished.

Confirmed

  • Reddit users criticized product evangelist Holly Li's demo for multiple glitches and weak example choices—flight and hotel booking can be done on Google in two minutes—and argued product management is dragging down engineering.
  • Blogger kimmonismus, comparing OpenAI Dot with xAI Grokbot, prefers the latter for its "Chief of Staff" setup and daily team standups.
  • markk called Dots "rushed": real potential, but the consecutive launches of Grok Bot and Meta Muse made it feel pushed out early, with plenty of bugs and an agent that feels bolted onto the ChatGPT app.
  • Cloud commentator QuinnyPig went from skeptic to "true believer" faster than expected, while admitting many complaints about the current state; nickbaumann adds the product is still early on its timeline.
  • cedricchee's hands-on take: Dots is not stronger than Codex but a new proactive-by-default AI collaboration paradigm—learning what you care about and acting on it—though latency is high and UX immature.
  • WIRED's side-by-side test of Dots and Meta Muse marks the official start of the personal AI agent war.

Why it matters

  • The rapid-fire releases of Dots, Grokbot and Meta Muse signal that the persistent personal agent race has officially begun; gaps in product form and delivery quality (demo failures, bugs, latency) will directly shape each player's reputation, and the reviewer consensus is a valuable direction with incomplete execution.