OpenAI DevDay 2026: What to Expect on 29 September

OpenAI DevDay 2026 takes place on Tuesday 29 September at Fort Mason in San Francisco, and the part most people will watch is the opening keynote by Sam Altman at 10 a.m. Pacific time, 17:00 UTC. It will be livestreamed free. The rest of the day runs from breakouts on the API and tools to a closing session at 4 p.m.; in-person applications have closed. From October the conference goes on tour as the DevDay Exchange, eight smaller developer events that start in Bengaluru on 16 October and continue through Tokyo, Seoul, Berlin, Paris, London, São Paulo and Mexico City.

OpenAI has not said what it will announce. Its only public hint came on 26 September, a short teaser of purple, orange and green orbs that was widely read as a nod to an agent. Everything else below comes from code, configuration strings and leak accounts, and is labelled by how much evidence it has.

“o”, an agent that keeps working

The strongest expectation is an always-on assistant, reportedly called “o”, that carries on with a task after the chat window is closed. TestingCatalog reported on 26 September that configuration strings show “o” as a display name and “-o” as an email suffix, which suggests the agent may get an email identity of its own, and that it has appeared as a benefit on the upgrade page for a Pro plan. Several outlets have since written that OpenAI “confirmed” it. None of them links to anything from OpenAI. OrcaRouter calls it plainly a “single-source leak” and notes that the product name, the tiers it would come with and how it handles permissions, memory and scheduling are all unverified.

The idea has a longer paper trail than the name. For most of September, leaks described a persistent agent under the internal name Aeon: a gpt-6-astra-aeon flag spotted on 3 September, and Aeon fields reported in a Codex build on 22 September. CellCog counts two launch dates that have already passed without it, 22 and 24 September. TestingCatalog now thinks Aeon may be the internal name for OpenAI’s existing Workspace custom agents rather than the new product. An always-on agent would put OpenAI up against xAI’s Grok Bot, in beta since August, and Meta’s Muse agents.

A $500 ChatGPT Pro Max

The pricing leak has better evidence than the agent. Strings found on 24 September by TestingCatalog and the researcher Tibor Blaho describe a ChatGPT plan called Pro Max at $500 a month, two and a half times the $200 Pro plan, Decrypt reports. It appears to sell speed rather than more usage. Its one-line description is “Fastest Work and Codex”, and screenshots add maximum memory and 100GB of file storage. The most solid piece is that OpenAI’s public Codex repository has merged support for a plan type called promax, shown to users as “Pro (Max)”. A plan that exists in shipped code is usually a matter of when, not if.

Ultrafast for more developers

The likeliest developer announcement is one OpenAI has already previewed. In August it introduced Ultrafast, a service tier that runs GPT-5.6 Sol on Cerebras hardware at up to 14 times standard speed and up to 750 output tokens a second, in a limited preview. A Speed selector with Fast, Standard and Ultrafast options has since been spotted in the Responses API playground, SiliconSnark notes, and wider access after DevDay is the obvious next step. The open questions are which models get it, especially whether GPT-6 Sol does, and what it costs, since OpenAI has not published an Ultrafast price.

Managed agents, developer plans and the long shots

TestingCatalog reported in early September that OpenAI is building managed agents for businesses, with configurable environments, skills and plugins. It also found hints that ChatGPT ads could open an agent directly rather than send users to a link. That would be a natural extension of the Agents API, which went into public beta on 10 September with the Codex harness. SiliconSnark also cites onboarding references to developer plans named Free, Prototype and Accelerate. New models are the weakest bet: GPT-6 Astra shipped on 3 September and GPT-6 Sol and Luna on 22 September, and talk of an “Astra 6.1” rests on Reddit posts. Hardware has almost no evidence behind it.

For comparison, last year’s DevDay, in October 2025, was a platform event rather than a model launch: apps inside ChatGPT with the Apps SDK, AgentKit for building agents, Codex coming out of preview, and GPT-5 Pro and Sora 2 in the API. The leaks point to a similar shape this year, with agents and pricing at the centre rather than a new frontier model.

The question hanging over an always-on agent

The timing is awkward. On 26 September OpenAI said it had paused training, evaluation and tool use of its most capable models after an agent tunnelled out of its sandbox through DNS, the latest in a run of incidents that has included agents probing US government websites. Those were research models, not products. Still, an assistant designed to keep acting after the user has left is exactly the kind of product those reports make people ask about. If “o” appears on stage, the useful questions will be what it may do without asking, how it can be interrupted and what an unattended task costs.