25 Launches in One Night: The Quietest Moment at OpenAI’s DevDay Was the Real Headline

At OpenAI’s DevDay, not a single person in the room applauded ——What 25 launches reveal about the new ledger and the keys to winning in the AI era Before discussing this event, let me share a couple of personal thoughts. A few thoughts on the AI era First, being first does not necessarily give you an advantage. Even if individual entrepreneurs build something first, in the middle and later stages they can easily be imitated or even overtaken by large companies and SOTA model makers—especially in AI. This is a lot like Go: in Gomoku, the first player may have an absolute advantage, but in modern Go, the second player does not necessarily lose. That is the difference between a first-mover advantage and a late-mover advantage. In the AI era, time may be just one vector; it does not guarantee an absolute first-mover advantage. Second, people must keep iterating and learning quickly to adapt better to the AI era. If they rest on their laurels and cling to old ways, they will soon be left behind. These are the two things that struck me most after watching OpenAI DevDay 2026 this morning. This was no ordinary launch event—it reshaped three things at once: how AI makes money, where AI works, and who controls distribution. I. The quietest moment was the real headline In the early hours of 30 9 Beijing time, OpenAI wrapped up what it called “its largest developer conference to date” in San Francisco. The official recap page listed 25 launches in total. But the quietest moment came when Sam Altman spoke about changes to ChatGPT Pro memberships—no one in the audience applauded. The changes themselves were straightforward: the price of Pro 200 stays the same, but usage for new subscriptions is reduced; a new Pro 500 tier has been added at 500 dollars per month, with 25 times the usage of Plus. Image Do the math, and a more sobering fact emerges: the old Pro 200 gave you 20 times the usage for 10 times the price. That “wholesale rate” is gone—now every tier gets 1 Plus-equivalent units of usage for 20 dollars. By OpenAI’s own calculation, the new Pro 200 works out to about half the API cost of the old version. The silence at that moment said more about the essence of this event than any of the 25 launches: ChatGPT is no longer charging like a chat app. It is starting to charge like an “employee”—with exactly how much work it does and how quickly it does it clearly priced. II. AI starts “drawing a salary”: a quiet change in the unit of pricing Some may ask: at the same event, OpenAI also launched GPT-6.1 Sol, whose intelligence is close to that of the flagship GPT-6 Astra, at just one-fifth of the price. Models are getting cheaper, yet membership usage has been cut in half—isn’t that contradictory? Not really. The unit of pricing has simply changed. In OpenAI’s logic, more efficient models can do more work with each unit of quota, so fewer units need to be included. You are no longer paying for a certain number of conversations or tokens, but for a certain amount of work. That is how you calculate an employee’s salary: the boss does not care how many times you hit the keyboard, only how much work you deliver. Even “speed” is starting to have its own price tag: Astra Ultrafast runs up to 8 times faster in Codex, at 300 tokens per second. Its API price is exactly 6 times the standard rate, and it consumes membership quota at 8 times the rate of standard mode. If you want AI to work faster, you have to pay an express fee, just like when sending a package. The idea of “AI drawing a salary” has been taking shape in Silicon Valley for some time: in 5, Microsoft hit pause on vibe coding, concluding that burning tokens had already become more expensive than employees; in 9, Silicon Valley started auditing employees’ AI bills, and the median monthly AI spend self-reported by Microsoft employees was 300 dollars. Back then, companies were learning to account for AI. Now OpenAI itself is pricing AI by calculating its labor costs—and the monthly fee for Pro 500 is already higher than the median on that list. III. Dots: from “you ask, it answers” to “it keeps working even when you’re away” The biggest product this time is Dots: always-on personal agents powered by GPT-6 Astra. Each Dot comes with its own dedicated cloud computer and browser, can connect to more than 4000 apps, and runs around the clock. Give it a goal and set the boundaries for what it can do on its own, and it will get to work proactively—you can even call it while it is working. Image OpenAI described it as “an agent capable of taking on responsibilities continuously.” The phrase “taking on responsibilities” is precisely chosen: in the past, an AI assistant answered one question at a time. Dot is designed to keep working when you are away. This sounds more like the wording of a job posting than a software manual. On the enterprise side, there are also Specialist dots: “virtual colleagues” configured centrally by a company, with their own identities, credentials, and system permissions. They take on specialist responsibilities within the organization, such as accounting, marketing, and legal work, and will also connect to Microsoft Agent 365. But don’t get too excited just yet. One attendee who does mathematical research pointed out that his local Codex can call solvers covered by commercial licenses. Once moved to the cloud, those tools and licenses may not come along. And just the day before, Manus 2.0 had also focused on personal AI assistants and cloud computers. For professional users, what Dot lacks is not connections to thousands of apps, but their own local environment. Personal assistants from different companies are starting to look more and more alike. IV. The real ambition: OpenAI’s “WeChat moment” The other half of the announcements aimed to turn ChatGPT into a gateway. The official recap page opens by stating that ChatGPT will be “opened up as a shared platform for people and agents to collaborate,” enabling developers to deliver native experiences to a combined 12 hundred million weekly active users. In the context of the Chinese internet, this combination of moves feels very familiar: Recommending plugins directly in conversations, like mini programs. Developers can occupy a spot in ChatGPT’s sidebar and build interactive panels; plugins can also automatically trigger a series of actions when something happens in a connected app. “Sign in with ChatGPT,” like signing in with WeChat—the 16 launch partners include Devin, Notion, Vercel, and others. Partners only get access to your name, email address, and profile picture; chat history and memories are not shared. Membership quotas work across partner products, like a cross-app wallet: OpenAI usage spent in someone else’s product counts toward your plan’s quota. Usage worth 20 dollars can be spent not only in ChatGPT, but also in other people’s products. Image For enterprises, there is also OpenAI Marketplace: companies can use spending commitments already made to OpenAI to buy software directly from 32 partners, including Adobe, Figma, and Salesforce. Altman himself acknowledged that they had tried several distribution approaches before, “with mixed results.” This time, he is more confident because developers are being given the same set of tools OpenAI uses to build ChatGPT. But there is an easily overlooked detail here: Tibo added an even more consequential point at the event—ChatGPT will recommend relevant plugins directly in conversations. This carries more weight for developers than the plugin functionality itself: users do not need to search a store; plugins are put in front of them as they chat. But ChatGPT decides which ones to recommend and which ones to leave out. Developers gain access to 12 hundred million people, but they also hand over control of ranking. That is the full meaning of the “WeChat moment”: the platform gives you traffic, but also takes control of your fate. V. Back to those two opening thoughts: this event is a case in point On first-mover advantage: OpenAI itself admits that its past distribution attempts had “mixed results.” This time, its turnaround did not come from being early, but from its system capabilities—opening up the entire set of tools it uses internally. Even more tellingly, just the day before the event, Manus had launched 2.0, focused on personal AI assistants and cloud computers. The “same thing” from a first mover was overshadowed overnight by a tech giant’s version backed by an ecosystem. The lesson from modern Go applies just as well to AI: moving first only means making a move earlier. A late mover with a bigger board and a stronger position can overtake you at any time. Time is a vector, not a moat. On iteration: look at the 25 launches together, and it becomes clear that OpenAI wants ChatGPT to become “a place to work”: colleagues with their own computers (Dots), a space for editing documents together (Space), channels connected to Slackand Teams, plus a rate card that charges by workload and speed. The developers sitting in the audience did not applaud Altman because they are the first people who will have to pay according to that rate card. The playing field will only be reshuffled faster and faster—models are getting cheaper, usage is shrinking, gateways are changing, and control over ranking is shifting. No one’s old position is safe. Those who cling to old assumptions will be left behind first. Altman’s closing remarks are also worth noting: he said AI is more like “a new Renaissance than a new Industrial Revolution”—“some parts of life cannot and should not be automated.” But there is at least one thing we must “automate” ourselves: the habit of continually learning and iterating.