OpenAI’s Dev Day drew 2,500 attendees and featured 20 launches, with the number potentially higher depending on how they are counted. The event has been held for four consecutive years. Figures presented onstage put ChatGPT usage at more than 1.2 billion people a week, while over 35 million people build with ChatGPT Work and Codex weekly.

The major product announcements included Dots, described as personal AI agents, ChatGPT Spaces and GPT-6.1 Sol, alongside other updates involving Sites, Codex and Work.

Oura has paused its planned IPO, but the reason has not been established. Market uncertainty was cited in a headline, while commentary questioned that explanation and suggested a gap between the price sellers sought and what investors would pay. One estimate put demand at 4.4 times the shares available; another account described orders at about four times the shares on offer. The offering was expected to raise as much as $2.2 billion and was described as primarily a secondary sale, not a necessary financing round: Oura planned to sell 13.5 million shares, while Forerunner and Lifeline planned to offer 36.5 million combined. Oura could therefore keep building without urgently raising cash, while those funds must wait to recycle their capital.

Oura was described as facing relatively little competition in smart rings, despite broader competition in wearables. Apple and Meta were raised as possible direct competitors, but no rumors were confirmed; Oura was also described as a potential fit for Apple’s ecosystem. AI-forward rings and devices with microphones are among the competition, though demand for microphones is uncertain. Commentary also raised the concern that adding one could make users feel the device was collecting too much data.

Long Lake has completed its acquisition of Amex Global Business Travel for $6.3 billion. The company has up to 30,000 employees; it was also reported to have never conducted a reduction in force and to be growing its headcount.

That growth was cited as an example of how effective AI use can accompany payroll growth, though no specific AI systems or causal link were identified.

AMD agreed to acquire AI research firm World Labs in an all-stock deal valued at about $8.2 billion. AMD said the acquisition will add leading AI research and development talent and help it build AI hardware, software and systems, particularly for physical AI and robotics.

World models could support synthetic data for training robots, while on-the-fly video game generation was also raised as a possible application. But one commentator questioned how soon the research could translate into substantial revenue, and noted that world models have not yet had a ChatGPT-like moment of broad adoption.

Sam Altman said he has mostly used OpenAI’s DOTS for work, where it has restored his attention and time and helped him explore new things.

He has not yet figured out how to use it for more fun in his personal life, but said he is inspired to try.

Sam Altman characterized AI’s development as a “jagged exponential” and predicted the next six months would be even more intense.

He said the technology had been capable for some time, but takes time to spread through society; people have used it much more over the past six months as the technology has also become more capable.

Sam Altman said capable AI models can do far more than routine tasks such as booking a haircut or coordinating plans with friends, including tackling major unsolved math problems. He said it will take time to discover the full range of what they can do.

Altman said more people are interested in making daily life less hectic than in what models can do for them with problems such as Navier–Stokes. He expects models to improve people’s lives in big and small ways that are difficult to imagine today, and called the focus on haircut bookings a “failure of imagination.”

OpenAI says it wants to build an open ecosystem where developers can bring apps into ChatGPT and users can take their subscription to other services and interfaces.

The company described this as a goal and said it plans to push portability further; it also expressed disappointment that some other companies restrict where their AI subscriptions can be used.

The possibility of AI agents messaging users first raises a risk of excessive notifications, Sam Altman said. He warned that companies could misuse frequent pings to drive engagement and expected some to get it wrong; he said OpenAI would try hard to avoid that.

Altman said his own experience with DOTS makes him want more messages, describing its current setting as too conservative. He also said the messages he has received so far have been things he wants, while acknowledging that proactive messaging could go wrong and that the right approach still needs to be worked out. A missed important email was offered as an example of when a user might wish for a reminder.

OpenAI was described as working in small, scrappy groups on difficult tasks under tight deadlines. A recently hired employee, who had previously worked at a large, relatively slow public company, reportedly framed the key to maintaining that kind of pace as increasing the work and expectations per person over time.

The argument was that too many people without enough meaningful work can compete over “fake” work, while keeping employees busy with work that has real impact may help large companies resist slowdown and internal politics.

Sam Altman said OpenAI works hard to fight AI-generated “slop” and has strong cultural norms against unnecessary expansion of text.

If someone uses ChatGPT to expand a text and another person then uses it to summarize the result into bullet points, Altman said, the better approach is to send the bullet points in the first place.

OpenAI’s chip program is set to start coming online in the first half of 2027 and scale substantially in the following years, a company representative said. OpenAI views the program as a way to build more computing capacity as demand grows among consumers, developers and enterprise customers.

The representative said he thinks the chips could give OpenAI a “huge inference advantage” for the watts available. He said the company will also need more efficient models and chips, as well as much more infrastructure, and predicted that people may want to run 10, 20 or even 100 DOTS.

OpenAI’s stated goal is to simplify model selection: rather than asking users to choose among models, a single model would make decisions for them and route requests appropriately.

The idea came up after DOTS was noted to have no model picker.

An OpenAI representative said he expects to be right about the rate of scientific discovery and believes people will care about much the same things and have similar motivations 10 years from now. He added that even with advanced intelligence, many parts of life will stay the same.

He said he tries to leave his mistaken predictions public and expects to get the order of developments wrong, citing haircuts, Hugging Face and Navier-Stokes as examples.

AI tools make building products fast and cheap enough that developers can try many ideas instead of committing all their effort to one. A developer with 30 ideas can now build them all, put them in front of users, and focus more on the ones that work; the advice to pick one idea was considered right two years ago, when building 30 was not feasible.

The approach favors many experiments, quick feedback, and acceptance that many attempts will fail. It also echoes venture investing: getting three bets out of 10 very right can, as one formulation put it, pay for the others.

AI alignment should not be treated only as an engineering problem, an OpenAI representative said. They called it dangerously wrong to claim that the scientific side of alignment has already been solved, and said more research progress is needed; they expect that progress to continue.

Engineering work is also needed, including better sandboxes and monitoring tools. The representative said that if models are to become much smarter than people, the science of alignment must be solved.

An outage occurred during OpenAI Dev Day demos.

The account linked it to the launch of “6.1,” which was immediately met with heavy demand, but noted that the precise circumstances had not been closely investigated.

OpenAI described DOTS as agents for work users delegate rather than tasks they must spell out and supervise step by step. A Dot can run on its own cloud computer, connect to a user’s computer, and appear in Slack with an identity linked to its owner. It can stay online, conduct proactive research, and adapt to feedback about the user’s goals and preferences, though getting value from it takes some initial time and guidance. Examples included finding files, updating a blog after a product change, organizing altered travel plans, and building games. OpenAI highlighted potential uses for solopreneurs and small-business owners, and said it plans to bring much of the same technology and capabilities to ChatGPT in the future, for what the speaker called its 1.2 billion users.

The same open-source Codex harness powers DOTS and OpenAI’s managed Agents API. OpenAI described continually adjusting the harness as models improve: it can compensate for current model limitations, then parts may need to be removed if they hold newer models back. Work on the harness focuses on efficiency, safety, and security. The product’s communication design also separates user-approved messages sent as the user from autonomous agent activity: after messages sent as users with an internal “sent by ChatGPT” label became noisy, the team moved toward distinct agent identities and names such as “owner-bot.”

OpenAI described specialist DOTS as agents for accelerating teams’ functional work, with their own computers and centrally governed identities rather than individual users’ identities.

It said it had an announcement that day about working with Microsoft on centrally governed identities. Inside OpenAI, DOTS run on Mac Minis and are provisioned like employees with identities and IAM roles; they also receive premium machines and handle tasks including procurement. Procurement and accounting were cited as early areas of success, with OpenAI expecting specialist DOTS to be useful in more functions and departments.

Space incorporates HTML visualization, making it possible to create interactive artifacts that users can click through, including a game embedded in a page. The speaker said AI-assisted HTML authoring with Dots is “basically free.”

They described conventional documents as designed for human authoring and somewhat inflexible, while also saying those tools are good. They characterized interactive HTML artifacts as part of a developing paradigm shift.

OpenAI described plugin extensions that let developers render custom interfaces and pages inside the ChatGPT app, giving them a way to reach ChatGPT users directly.

A meetings plugin launched that day was cited as an example of an external integration. OpenAI also said connectors can make integrations faster and improve the user experience, while custom rules let companies limit what agents can do with tools.

One view is that AI interfaces should match users’ preferences and meet them across modalities, rather than settling on a single mode. A product manager described using Slack during the workday and talking to a DOT by voice while commuting. The voice experience emerged when teams combined existing capabilities, rather than being designed from scratch as a new product concept.

Discoverability remains a challenge: a retail store in Charleston teaches visitors to use Codex and ChatGPT, while people outside tech may not realize they can ask AI whether it can do a task. As an example of active intelligence, DOTS might notice an urgent request for a headshot, find it, and ask whether to send it. This was presented as a possible capability, alongside a prediction of a step-function increase in the value people get from existing models.

OpenAI said it is combining Codex’s agentic capabilities with ChatGPT. The company said it rewrote chatgpt.com in five weeks to provide the Codex app experience, which is now available on the web as the ChatGPT desktop experience.

ChatGPT work on desktop now runs in the cloud while retaining access to a user’s local machine. Codex Cloud launched that day, and OpenAI said it was working on further “toggle merges” that month.

OpenAI says it aims to put new capabilities into users’ hands even when product integrations are not complete. Codex was one example: the company said its capabilities needed to reach everyone even though it could not yet bring chat and work together. OpenAI said it is taking a similar approach with Space and Dots.

The company said it extensively dogfoods its products, with thousands of employees relying on them for every part of their jobs, and uses that experience as a feedback loop. It described the products as works in progress that are changing quickly, and said it receives frequent user feedback, including many tags on X.

OpenAI described many of its products released that day as “take 5” versions of earlier internal work. Dots went through many variations, including local and cloud versions, and Space through multiple iterations. The team had previously offered a chat feature called Canvas and developed internal knowledge-based products.

Codex Cloud was described as a “full 2.0”: Codex began as a cloud product, moved into a period of local agents, and is now returning to running a full local agent in cloud environments. The OpenAI team member said rapidly changing models make it difficult to time products for when a model is best suited to them; some efforts can be slightly early or “slightly right.” They added that the products available that day felt “magical.”

OpenAI describes Space as a way for teammates to collaborate across work that is otherwise spread among apps, documents and Slack, and as an extension of ChatGPT bringing Codex and ChatGPT into a more collaborative, multiplayer setting. The company plans further work on the product. One example is an interactive chart created in a shared page: a Dot can research connected data sources, build a small app and share it with a team. OpenAI says it has worked on the challenge of making such experiences viable, since an instant chat model cannot write a full web app for every request.

The exchange also linked faster models to enterprise workflows. A speaker referred uncertainly to “8 times faster” as ultrafast; the exact benchmark was unclear. The speed was described as especially useful for dynamic interfaces, while pricing drew some negative reaction and Tyler’s response was that “time is money.” The discussion raised whether users should manage speed and workload trade-offs or whether resource allocation should be abstracted; it was described as being abstracted quickly.

An internal OpenAI chart showed engineering and research using coding agents, with other groups quickly following, according to an OpenAI employee.

The employee cited collaborative tasks such as pulling data, researching in Slack and planning events.

Peter Steinberger traced OpenClaw to frustration with slow or ineffective computer tasks and the inability to simply talk to a computer. He said it showed a broader use for AI than summarizing email: with the right tools, it could install software, make purchases or take other actions on a computer, and could be proactive. He also described its presentation as unusual and more human-like.

Steinberger argued that existing workflows are already becoming outdated as models advance. He said Loona was now “too cheap to matter” and UltraFast let him avoid 50 terminal windows and stay focused, adding that builders cannot keep up. He acknowledged that hype around OpenClaw inflated expectations beyond what it could do: models and software still had limitations, and the product remained hard to use. Users repeatedly rewrote it. The OpenClaw moment was also described as feeling like a renewed burst of the town’s hacker culture, alongside a remark about people returning to hardware.

Peter Steinberger said it is still unclear whether AI will center on one personal agent or a team of agents. He leans toward the personal-agent approach but sees both as possible, and said the industry has not yet worked out how personal and team agents would interact.

He described today’s chat-and-work interfaces as a compromise, noting that introducing new ways of interacting can confuse or annoy users. One possible architecture, not a settled plan, would have a personal agent delegate a complex task to a manager agent, which could then coordinate smaller subagents. Steinberger said he is exploring many ideas.

Peter Steinberger said he moved his team into a shared space to work with agents. After initial pushback and a day of embarrassment about how privately people used them, the team began learning from one another’s different approaches.

The exchange also raised the value of sharing AI-assisted research prompts and results: Tyler Cowen’s economic research was cited as an example of work others might want to inspect, since they may not think to ask the same question. There is not yet an established habit of sharing ChatGPT transcript links, though Groq responses occasionally appear. The discussion of attribution emphasized that the person who thought of the question deserves credit even if AI did the research, while noting that people may take credit for the result.

Retailers’ efforts to block bots can also prevent shoppers’ personal agents from using their systems. For scarce, highly sought-after sneakers, a race to click first may no longer be a fair way to allocate purchases. A cited Steam Machine example used a week-long registration window followed by a lottery, giving people the same chance whether an agent clicked in 0.1 seconds or a person clicked an hour later. Some services may need to rethink rules built around human shoppers; otherwise, users may simply buy elsewhere.

In a possible future, agents on both sides of a transaction could check inventory and sizing behind the scenes, much as credit cards simplified the mechanics of payment. Agent competition could then play out largely out of users’ view.

AI agents, including OpenClaw and ChatGPT, are largely user-prompted today, but a more proactive model could keep analyzing context and prepare help before a request. One possible use would be checking a calendar for an upcoming interview and assembling research for it; this was an example, not a product plan.

Always-on agents would pose a hardware challenge as well as a design challenge: the approach could require substantial inference computing and more chips.

Many people can now try AI agents, a development viewed positively.

But the people building highly intelligent tools are not good at explaining what users can actually do with them. Much work remains.

A proposed AI agent would break a large request into a hierarchy of tasks while accounting for the user’s intent and desired timeline. Rewriting Postgres in Rust was offered as an example: the agent would need to establish whether the request was serious, plan the work, and determine whether the result was needed immediately or in a couple of months. This was described as a design idea, not a released product.

The broader view is that AI-generated code is still “slop,” but will improve as models do. Trust in models has grown despite earlier mockery, and coding habits are shifting from inspecting every line to scrolling through code; related workflows may need rethinking.

Codex can be set up for computer use and given conversational instructions to carry out tasks on a PC. In another example, an agent read and compiled the Linux kernel, after which a webcam worked.

Codex also fixed the display driver on an ESP device overnight, with a webcam aimed at its screen so the agent could observe the result. An attempted audio fix took much of the night: the person kept waking and calling “hello ESP” while the agent tried to work out that the volume was turned down; it eventually learned.

These examples point to a broader possibility: an agent could become an operating-system-like way to interact with a computer. Building a personal OS remains difficult, but advances in models and access to enough tokens could make the idea less far-fetched in the future.

Work spanning roughly two teams includes substantial work on the foundation side.

Inside OpenAI, new ideas are tried and some are considered for development into first-party products; ideas also flow back the other way. Not every approach transfers across different scales, since building for each requires a different method.

People one attendee spoke with at OpenAI Dev Day seemed especially excited about the developer ecosystem and OpenAI opening up Sign in with ChatGPT. That focus surprised the observer, who had expected people to be most excited about other things. The announcement was called a standout.

The developer appeal was framed as letting builders focus on products for customers rather than on making token economics work.

An investor who works with “70-some” portfolio companies says they usually advise separating a fundraising launch and other announcements, rather than launching them on the same day, and spreading them over two weeks.

After seeing Dev Day live, the investor said bundling many announcements could make sense: having them land together creates a moment the investor finds fun and puts pressure on the team to deliver.

OpenAI’s computer-use speed has increased tenfold over one year, according to a speaker, though no measurement method was specified. The improvement was attributed to both the model and its harness: the model needs less time to choose a reliable next action, while the harness can run operations in parallel and reduce delays. A separate auto-review agent, internally called Guardian, watches the primary agent and can intervene if it is about to enter sensitive information on a questionable domain.

Astra’s ultrafast mode was cited at 300 tokens per second. Separately, the speaker estimated about 20 tokens per action and roughly 30 actions per second; how those estimates relate to the 300-token rate was not clarified. The speaker also forecast real-time computer actions within the next year and said a standard OpenAI model with computer use could beat the model from the company’s first Dota competition, whose 10-year anniversary falls next year. The comparison was described as potentially close because actions per minute matter alongside reasoning.

OpenAI says it is working toward a system that would let users describe a goal, deadline and budget in natural language, while the system configures the resources needed to meet them.

Its DOTS product already has no model picker, is available around the clock and varies its effort depending on the request; OpenAI cautions that it may not be perfect.

An OpenAI team member says they combine conversations with promising candidates with discussions of ideas the team is pursuing or considering speculatively. The conversations help them assess how a candidate might approach the work as part of the team, while also developing their own thinking about it.

They say interacting with people and looking for new candidates takes about 20% of their time.

OpenAI is considering a shared memory model for DOTS and agents more broadly: a system maintained across an organization that combines individual notes with collective ones. For example, an organization could note that it is a Rails shop and what an agent needs to know; the agent could then read the relevant material.

OpenAI also says it is continuing to develop its memory system so it can preserve coherent memories over very long periods, retaining feedback rather than forgetting it after a couple of weeks.

ChatGPT Sites reached 10 million sites created in under three months since launch.

One use case combined Codex-generated remodel visualizations for Zillow listings with neighborhood property comparisons, itemized estimates for architects and materials, and a comparison of projected costs with potential value over a 10-year holding period. The person who built it predicted that tools like these could accelerate economic activity by encouraging more remodeling and purchases as people gain confidence.

ChatGPT Sites launched plugins today, enabling users to connect calendars to interfaces they build.

An OpenAI product manager described using the feature to create a family organizer linked to Google Calendar, including children’s calendars, instead of buying a dedicated family calendar product such as Skylight.

ChatGPT Sites lets users share sites and add co-editors and co-authors. OpenAI’s IT team uses Sites to share information about employee locations with colleagues managing on-the-ground security, including arrivals, events and weather. In another use case, Deep Research generated potential guest suggestions, which were put into a site with thumbs-up and thumbs-down controls for team review and sharing.

OpenAI plans closer links between Sites and Pages: each will be able to reference the other. Slides are expected to launch soon and will be easy to add to Spaces. OpenAI is interested in making site creation available to everyone, including free ChatGPT users, but is still working through infrastructure and pricing. Ad-supported access was mentioned as a possibility, not a settled model; no availability timeline was given.

A cited milestone was 10 million sites. A forecast put the total at 10 billion “soon” and predicted more websites per person, without specifying a timeframe.

ModRetro made 8,000 transparent custom Chromatic devices for the event so every attendee could take one home, according to a company representative. The device had been announced on stage by Sam about an hour and a half earlier; one was then listed on eBay for $4,000. The listing was discussed as a possible resale opportunity, but no sale was confirmed.

ModRetro also built a custom plugin for Codex that lets users vibe-code games and flash them directly onto a Chromatic cartridge, rather than playing only in a laptop emulator. As an example, CEO Torin and his five-year-old created a game that Torin flashed to a cartridge so his child could play it right away.

ModRetro released the M64, its Nintendo 64 console, about a month and a half ago. A company representative said it is doing “incredibly well,” without specifying a measure. Much of the feedback, the representative said, came from older people introducing the console to their children; the representative added that he was 15 when the original Nintendo 64 came out.

The representative said the console’s physical constraints can encourage parents to sit and play Mario Kart with their children. He contrasted that shared experience with a child alone on an iPad playing Roblox, describing the M64 experience as social and companionable.

Dedicated devices can create a stronger emotional bond than all-purpose technology, while hardware itself may change more over the next 10 years than it did in the previous decade. A yellow Sony Walkman that could get wet was recalled as a prized childhood possession; Spotify was considered technically better but less emotionally meaningful. A point-and-shoot camera and the iPod Shuffle were other examples, with the Shuffle’s unpredictable song choices making listening feel like a journey. One view was that much of what is good in the physical world has not changed in 60 years, with machine intelligence singled out as a meaningful exception.

New software paradigms were cited as one possible driver of hardware changes. Peter from OpenClaw was described as frustrated with existing operating systems and as saying some tasks could make more sense with an operating system generated almost on the fly. That was an idea, not an announced product or confirmed plan.

Over the past five years, more people have been seen using 10-year-old point-and-shoot cameras, despite their lower image quality compared with phones. One 35mm Olympus compact, which originally sold for $150 in the 2000s and was discontinued about 15 years ago, has become much harder to find: its owner recalls buying replacements on eBay for $50, then seeing the price rise to $100, $200 and, when shopping for one as a birthday gift, $400. The cameras are now almost unavailable, according to the owner.

The price figures are personal observations, not broad market data. The owner suggested that using an older camera can signal something about its user and reflects appreciation for the devices beyond their technical specifications.

A critique of connected TVs focused on software that can become less responsive over time, forcing users to click repeatedly to navigate.

The criticism was qualified by the possibility that some slowdown could be related to not updating the software.

Two hackers’ YouTube analysis reportedly found that microphones in the most modern LG TVs were always listening, storing audio locally while the TVs were disconnected from the network and uploading it when they reconnected.

The supplied material does not establish which models or software versions were affected, or independently verify the claim.

One Meta Quest user said getting the headset ready to play a game took about 15 minutes and involved seven passwords or codes, including credentials for Meta, Instagram, Facebook and Oculus VR.

Separately, downloading Netflix on a Samsung TV required signing in to a Samsung account. The account demands were criticized as not improving the customer experience, with the view that companies want users’ data to sell more ads.

Lower production costs and AI tools could help individual creators make more games, software and video, while increasing both low-quality output and the chance of standout work. One YouTube creator recalled having 280,000 subscribers after seven years on the platform, then reaching 10 million 18 months after starting a daily vlog. The creator attributed that inflection partly to more accessible production, cheaper cameras and faster editing enabled by technology.

Another creator said AI can help people get past technical blockers faster, potentially letting one person do 10 times as much in the same period. The view was that a wider opening for creation will produce more poor work as well as more quality work; a fully AI-generated film that is undeniably good was also predicted, not reported as an existing development.

As AI-generated video becomes more common, a central question is whether someone has checked it. One example is an Instagram account posting probably 20 sim-racing tutorial videos a day; the advice may seem generally right, but material drawn from online forums can be outdated. That makes human review and editorial curation important to accuracy and trust.

AI-generated videos can still work for viewers when they are judged factually correct and have human editorial oversight. Meanwhile, creators such as Mark Felton and H.I. Sutton show the appeal of personal expertise and narration. One forecast suggests AI content could become roughly 10,000 times more abundant in probably two years, while leaving room for expert human creators.