OpenAI DevDay 2026 keynote: What worked and what broke
Dottie stalled, voice control failed again, and typed commands changed a 3D venue in seconds. Follow the live demos, Lavender the robot duck and the worldwide reset with its scope still unclear.
OpenAI's first live agent demo got stuck waiting. The OpenAI DevDay 2026 keynote got more revealing when the presenters had to recover.
Later, voice control failed again, a typed command changed a 3D venue in seconds, and a robot duck quacked on cue. The event ran on 29 September in San Francisco.
The approximate minute markers below follow the supplied transcript from Sam Altman's opening; they aren't exact timestamps for every version of the video. Simon Willison's live blog provides a separate eyewitness check on the important moments.
OpenAI DevDay 2026 keynote: what happened on stage
Keep an eye on what was announced, what actually ran on screen and what was promised for later. The keynote moved quickly between all of them.
| Approximate minute | Moment | What the audience actually saw |
|---|---|---|
| 0 to 2 | Welcome and dots announcement | A recap of requested features, then the new ongoing-agent product. |
| 3 to 7 | dots introduction | A promotional video and explanations of delegation, cloud computers and connected apps. |
| 8 to 16 | Space and Holly Li's demo | Shared documents and team handoffs, a stalled voice exchange and an acknowledged coding error. |
| 19 to 22 | Sol, Ultrafast and plans | GPT-6.1 Sol, a side-by-side rocket example, Pro 500 and reopened Pro 200 subscriptions. |
| 23 to 26 | Research segment | Altman's research intern claim and Tejal Patwardhan's internal research figures. |
| 28 to 30 | Developer infrastructure | Cloud Codex, Security Cloud, Agents API, AWS agents and privacy announcements. |
| 32 to 35 | Romain Huet's 3D venue | Voice trouble, followed by typed changes to the virtual venue. |
| 35 to 36 | Free-ticket app | Live coding a way to select attendees for next year's event. |
| 37 to 40 | Computer use and cloud work | Playing a generated game, inspecting an app in a simulator and starting a Rust rewrite. |
| 41 to 43 | Lavender | A robot duck, voice interaction and an audience drawing. |
| 44 to 47 | Distribution | Sign in with ChatGPT, plugin extensions, Sites and Marketplace. |
| 51 to 52 | Red reset button | A countdown and a stage announcement that the usage reset was worldwide. |
Minutes 9 to 16: Dottie's slow morning
Holly Li used a fictional music app called Blossom. Her dot, Dottie, was meant to gather launch context, feedback and work delegated through connected tools. The appeal was easy to see: you spend less time piecing together what happened overnight and more time deciding what to do about it.
Around minute 11, Li asked for an update on user testing, and Dottie said it was checking. The wait kept going. Willison recorded the phrase "Dottie is having a slow morning" and the reply "still checking" in his live notes.
Li moved on to Space, showing shared documents, interactive feedback views, mentions and team handoffs. You could see the workflow there: information collected into a page where people and agents worked together. That was easier to assess than the agent's personality.
Around minute 16, Li acknowledged an error affecting the earlier Codex thread. Dottie was supposed to build the app using Codex on a laptop and open it in a simulator. The demo showed the intended workflow without confirming that this delegated build finished correctly on stage.
Waiting is part of using an ongoing agent, and a slow task can still be useful. But the status needs to tell you what's happening. If the assistant keeps saying it's checking, you're left deciding whether to wait, repeat the request or step in.
Minutes 19 to 26: speed claims and the research intern
The model segment went more smoothly. OpenAI presented GPT-6.1 Sol as a cheaper option close to Astra's capability, then showed Ultrafast with agents building rockets side by side. The faster version launched while the other was still building.
The rocket example made speed easy to see, but it wasn't a controlled benchmark of every coding workload. The official recap advertises up to eight times faster token generation in Codex, reaching 300 tokens per second, and up to six times faster in the API. Your completed task can still wait on tools, tests or a human decision.
Altman then said OpenAI had reached its AI research intern goal. Patwardhan described internal research use, including more than a third of day-long tasks completed without intervention as of July. Those were OpenAI's measurements, presented on stage rather than independently assessed during the keynote.
In a fast sequence of launches, a rocket animation can blur into an internal evaluation chart and then a product access announcement. Each answers a different question. None alone tells you what an agent will reliably finish in your repository.
Minutes 32 to 36: the venue worked after voice failed
Romain Huet entered with a camera feed and used Ultrafast to interact with a 3D representation of the venue. Voice input didn't cooperate, so he switched to typing.
Typing worked. Willison reported that Huet asked to put the livestream on the virtual venue's big screen, and it appeared shortly afterward. Watching the scene change made the speed claim easier to grasp than a tokens-per-second label.
Both parts deserve to stay in the recap: voice control was unreliable in that moment, and typed editing was responsive. Calling the whole sequence a failure loses the useful result. Calling it an uninterrupted voice demo loses what broke.
Huet then live coded a ticket giveaway app. Around minutes 35 to 36, the request changed from selecting three attendees to six for free DevDay 2027 tickets. A giveaway tied the coding demo to the event and gave the audience something more relevant than a sample dashboard.
You saw a small app under stage conditions and a short path from an idea to something the audience could use. That didn't establish production readiness or an independent audit of every selection mechanism. Keep that limit in mind when judging the demo's success.
Minutes 37 to 43: a Rust handoff and a robot duck
Astra next used computer controls to play a newly generated game. A remote app workflow showed an iPhone simulator inside the Codex interface, and Huet sent cloud Codex a request to rewrite the backend in Rust. Agent work here covered both interacting with software and handing off a coding job.
The keynote showed the rewrite starting, then moved on without waiting for a completed migration, test results or a deployment. You could see how to hand off longer work while continuing with something else. You couldn't assess the finished migration from that sequence.
Lavender, the robot duck, was the most memorable physical prop, connected to voice and image capabilities. Huet asked it to look at the audience and make a drawing. It quacked during the exchange, and the generated picture appeared on screen.
Willison identified the hardware as a Hugging Face Microduck robot. The demo showed a developer connecting capabilities to hardware that responded to its surroundings. It wasn't an announcement of an OpenAI consumer robot.
Hacker News noticed the wait and the speed
The reactions were mixed. Individual posts tell you more than a blanket verdict that everyone loved or hated the event.
On Hacker News, jdw64 singled out the slow dots demonstration. In the separate Sol discussion, vb-8448 was excited by Astra's advertised 300-token-per-second Ultrafast speed. The wait bothered one commenter; the prospect of faster output excited the other.
On X, OpenAI's Thibault Sottiaux attributed the live demo problems to rolling out updates simultaneously. His post, readable through Twiscan, also said the team intended to keep doing live demonstrations. Treat that as the company's explanation, rather than an independently established diagnosis of each failure.
Faster models looked useful, and the assistants still needed the surrounding system to work. The live demos exposed that gap in a way a recorded product video can't. The reactions picked up on both sides of it.
Minutes 51 to 52: what did the red button reset?
The closing countdown ended with a physical red button. Altman said the reset applied worldwide, and Willison confirmed that announcement. It wasn't described as a perk only for people in the room.
As of 30 September, we found no official post defining which usage counters were reset, how it affected every plan or whether it covered API billing. Without that scope, you can't infer a specific credit balance from the stage announcement.
The reset also doesn't reverse the written Pro 200 allowance changes: the Pro help page still describes the transition for eligible subscribers. Check your usage dashboard to see the effect on your account.
FAQ
Where can I watch the OpenAI DevDay 2026 keynote?
The keynote recording is on YouTube. Use the minute markers here as approximate offsets from the opening, rather than guaranteed player timestamps.
Did the dots live demo fail?
The voice update stalled, and Li later acknowledged a Codex thread error. The shared-work demonstration continued through other parts of the workflow.
Was Lavender an OpenAI robot launch?
No. Lavender demonstrated voice interaction and image generation connected to Hugging Face Microduck hardware.
Did OpenAI reset everyone's usage limits?
The stage announcement said the reset was worldwide. Check your account for the effect on your limits; the scope caveat above means you shouldn't assume a particular credit grant.