The Key Elements of OpenAI's DevDay

Jiang Wu · 2025-10-08

*Originally published in Chinese on October 8, 2025.*

OpenAI's annual developer conference, DevDay 2025, took place at 1 a.m. Beijing time on October 7, 2025 (October 6 in San Francisco). It showed a complete toolchain from app integration through agent development to code assistance, which marks a strategic shift from model provider to AI operating system. The main announcements:

- Sora 2 API: the API for the latest video generation model, which can be integrated into apps with a simple call, supporting custom video length, resolution, and style.

- GPT-5 Pro API: a new-generation model interface opened to API users, with noticeably faster reasoning, especially suited to hard tasks that need high accuracy and deep reasoning.

- Realtime Mini API: real-time interaction in a smaller, cheaper voice model, with costs 70% lower than the earlier version.

- gpt-image model: an image generation API. With this, OpenAI has a complete API matrix covering text, image, video, and real-time interaction.

The logic of rebuilding the app store directory

If it rebuilds the app store directory, it creates new choices and new cognitive costs for users, which is a bit of an unnecessary extra. The more central element is for the model to identify the user's precise need.

Managing input and output for services in a unified way keeps delivery stable. I think that may be a big trend ahead.

An app directory only hands the right to choose, and the uncertainty, to the user, and asks users to tolerate what the platform can do. In the long run, an integrated service capability should stress standardized input and output services and should not have too many steps and procedures in the middle to match forms. This transitional state will keep evolving.

The value of a workflow-building tool

From what I can see, the final aim of this workflow-building product may not be to encourage developers. It is more about freeing ordinary users' creative imagination. The platform's base capabilities are limited, and so is the set of scenarios it can make concrete.

Only when it opens creation to most users, lets them build new workflows from their daily scenarios, and finally share them as templates, does the logic of a system emerge.

For most productivity tools and platforms today, for example, one potential business is selling templated services. An orderly kind of disorder is a way of building that meets many scenario needs. The logic of the disorder is to have some orderly people build the tools and structure the matching scenarios. Within that principle, the ecosystem of software services and the platform's value are maximized. Another principle in it is whether the other tool platforms cooperate.

Connecting the data systems

What it hasn't solved yet is the data system. On service capability and basic workflows, I believe the problem won't be big in the short term. In the long term it depends on whether it can connect the user's data center into integrated data service. Today, some vendors will plug in when they need its traffic. Over a longer stretch, what barrier it will use to complete the service is unknown for now.

There is also a new problem. If it wants to be the user's data center, or at some level the core organizer of a data system, it faces major decisions on how to structure and group user data. Storing memory alone is just an entry point today. In a user's data scenarios, memory is one core element, and the more important thing is to use memory in the right scenario. I think that may be a major challenge.

Behind the AI OS logic is a re-sorting of supply relationships

Everyone thinks of Sam Altman building an AI OS, and in essence he should have that idea and expectation. But the core logic of an AI OS lies in rebuilding the core framework and carrier through which users connect to this system's services, and a new system underneath is secondary. Today it runs as software on different platforms. He also bought a hardware team earlier, so he will certainly make hardware.

What form will the hardware take? Will it be a complete device? I think everything is unknown today. But we can try to sort out what the core value points behind this logic might be.

Hardware as a data input center

In essence hardware replaces the user's data input center. On that view, what it wants to take is the full input from the user. Whether output is needed is not certain today. Since the system's dynamic capability is still limited, my guess is that the hardware is meant to offer a structured way of input and output, with all processing done on the model side.

The capabilities in this model system:

For now a wearable device seems to match its expected scenario: it only has to solve the input end, and output can use the devices users already have. So at present the core thing it wants is the user's input.

Data storage and usage strategy

The second value point is how it stores and uses user data. Does it store data in the cloud under a personal account system, or extract it by the user's feature values and store it by scenario? I don't think this is particularly hard technically. The harder part is recognizing and confirming the user's scenario.

A pile of hardware companies are trying to solve this today, but might it be something a hardware company should not be solving at all? It looks more like a scenario capability inside a service system, or something for a vertical small model to solve. I have no answer for now. If I'm optimistic, they will ship a solution by the end of this year, or they will buy a similar company.

Service APIs and the required basic modules

OpenAI basically strengthened every capability it could open up in its base capabilities this time, but for it, this is far from enough. Beyond the part of the base capability system that users can see, there is much users can't see, and they are researching it in depth. The reason it doesn't dare release carelessly today is that it has to consider what the whole partner system thinks of its ecosystem's value. So it picks some ecosystem partners to fill in and reinforce its base capabilities. But for modules that are too basic, it will certainly go deeper in the future. Whether text, audio and video, or anything else, the company will find ways to meet the need, because only after it solves those needs does it have a more solid commercial value. The typical goal is to serve users' needs, and broken apart, these needs are a very large business, whether to B or to C. I think they are very firm about this.

OpenAI will be a cloud service company in the future

A conclusion I can't support with data. Let's wait and see.

Home · All articles · Topics · 简体中文