OpenAI DevDay Exchange Tokyo Recap: Practical Uses of dots and Codex

On October 20, 2026, OpenAI held the developer event "DevDay Exchange Tokyo" at Bellesalle Shinjuku Grand. More than 500 people attended, including developers, researchers, and students. OpenAI's speakers introduced products and technologies and how to use them, based on the announcements made at DevDay held in San Francisco at the end of last month.

Focusing on the keynote, we report on the use of dots and Codex, and the mechanisms that connect ChatGPT with external services.

In Japan, dots' weekly active users rank second in the world

OpenAI Japan Representative Director and President Tadao Nagasaki took the stage for the opening. He explained the purpose of the event to Japan's developer community and also showed the spread of usage within the country.

OpenAI Japan Representative Director and President Tadao Nagasaki, who took the stage for the opening

According to the data OpenAI presented this time, Japan ranks second in the world for dots' weekly active users, and is the largest in the Asia-Pacific region. Nearly 60% of Japan's weekly active dots users used it on three or more days in the same week, and more than 60% also used Spaces or Pages that week.

In enterprise services, the number of contracted seats (user licenses) for ChatGPT Enterprise in Japan more than doubled year on year. The combined number of contracted seats for ChatGPT Business and Enterprise was nearly double year on year. API usage is also spreading, with the number of active OpenAI API accounts increasing by more than 60% compared with the same period last year.

Research and development that dots advances in the cloud

Next, Vijaye Raji, CTO of OpenAI's applications division, took the stage and explained how to use dots, Spaces, and other tools for work. Details of the products and technologies announced at DevDay in San Francisco are summarized in last month's report.

Vijaye Raji introducing the announcements from DevDay

dots is an AI agent that, based on the user's goals and past interactions, advances research, document creation, software development, and other work in the cloud. In addition to one-off requests, it can also be entrusted with scheduled execution at specified dates and times, or with work triggered by updates to supported connected services. Even during work, users can consult via messages or voice calls, add instructions, or change direction.[1]

Raji cited as a feature of dots that it continuously grasps the status of work and notifies users of issues they had overlooked or proposes next steps. The official guide also explains that it proactively investigates information it has read permission for, and makes proposals and confirmations by linking it to past work. For example, it reports schedules or asks about the progress of work discussed previously.

In a slide from the presentation, he also showed a use case for proposing budget allocation among business divisions for the next year's investment plan.

Examples of requests to dots, from budget allocation proposals to bug fixes

As an example of use in development, Raji took up responding to bug reports posted on Slack. dots investigates the cause, fixes the code, runs tests, and creates a pull request.

Another example of use in development: requesting migration from an API scheduled for termination and preparing a pull request for review

dots can also use data and tools from connected apps and the user's local PC. Work in the cloud continues even after the PC is closed, but work that uses the local PC requires keeping that PC online and the ChatGPT app open.

Raji introduced Specialist dots for accounting, email marketing, legal analysis, inquiry handling, and other tasks as a preview for enterprise customers.

In his presentation, Raji also took up ChatGPT Spaces as a place where people and agents share pages and files to work together. Research results, plans, and created materials can be gathered in the same place and checked and revised together with team members and agents.

As specific examples, he cited use cases such as saving meeting notes created with the meeting plugin to a shared page, and creating and editing presentation materials. He further discussed operations for continuously researching information and keeping page content up to date [2].

Agent development supported by GPT-6.1 Sol and the Agents API

To run agents continuously, response speed and cost are important in addition to inference quality. In his talk, Raji introduced GPT-6.1 Sol and the high-speed output Ultrafast mode [3].

GPT-6.1 Sol is a model that aims to combine performance close to GPT-6 Astra with low cost for complex coding and business tasks. According to the official pricing and specifications, under standard API pricing, the per-token prices for input and output are set at one-fifth of Astra's. [4]

In a demo in which a 3D rocket was created and launched, Raji compared the speed of the normal mode and Ultrafast mode. While the normal mode was still working on the creation, in Ultrafast mode the rocket on screen was completed and launched.

For developers, he also introduced the Decisions API, which processes classification and judgments with low latency, and the Agents API, which embeds agent capabilities into custom apps using Codex's execution infrastructure [5].

Mechanisms for continuing work with Codex and the division of roles among agents

Bonnie Chatterjee, who leads Applied AI Engineering (API) at OpenAI, gave a talk titled "What Comes After Loops?" explaining mechanisms for agents to carry out long-running work and approaches for combining multiple agents.

The foundation for this is the Codex harness. It manages the instructions and conversation history passed to the model and the results of tool executions, and executes the tools the model requests. The model decides its next action based on the returned results, and repeats this exchange to advance research, implementation, and verification.

In long-running work, loaded files and tool execution results accumulate, approaching the upper limit of the amount of information the model can handle at once. Compaction (context compression) is therefore used to reduce the amount of information passed to the model while retaining the information needed to continue the work.

Structure of the Codex harness. Work proceeds by combining interaction with the model, tool execution, and context compression.

Even if the necessary information can be retained, processing may stop when interim results are reported while the work is still incomplete. To address this issue, Chatterjee introduced Codex's /goal. By setting a goal in advance, Codex can check achievement status even after returning an interim report, and continue working if the goal has not been met. The slide showed an example of investigating the cause of slow search, trying a fix, and verifying the effect.

Example of requesting search speed improvement with /goal and proceeding to analysis, fix, and verification

The official guide on /goal describes usage in which, in addition to the goal, the verification method and scope of work are defined. Work does not continue indefinitely; it stops when the configured budget limit is reached, when the user pauses it, or when it can no longer proceed without user input.

Through such mechanisms, even a single agent can carry out long-running work. When work is divided among multiple agents, coordinating roles and duplicate work incur costs. For this reason, Chatterjee presented the idea of first clarifying the goal and verification method and starting with a single agent. Dividing work among multiple agents is considered when there are tasks that can proceed independently, such as large-scale refactoring or game production.

Chatterjee also explained that when dividing work, the deliverables, scope of responsibility, and shared constraints of each agent should be defined. Within that scope, each agent should be able to make judgments and proceed with the work. Relevant agents coordinate directly with each other to avoid duplicate edits and excessive intermediation. However, they are instructed to report and consult when changes to deliverables or scope of responsibility are needed, or when a problem cannot be resolved.

To align the understanding of each person in charge, a shared work log is also used. Chatterjee showed on a slide an example of AGENTS.md that records instructions to agents.

Create one work log that records the goals, current instructions, person in charge, incomplete tasks, and decisions. New instructions should be checked against work already in progress, and relevant items should be updated. Record decisions and their outcomes, and notify the relevant parties individually and directly. After work is completed or handed off, read the work log and confirm the current state before continuing the work.

In the game production example, tasks such as level placement and cooperative play are divided among the team, and goals and decisions are compiled in a shared log.

Task division and shared logs in game production

Chatterjee also presented a configuration example in which models are used differently depending on the task. In that example, an agent using Sol plays the central role, asks Astra for advice on difficult problems, and assigns routine work to the lightweight model Luna.

As an example of large-scale coordination, the talk also touched on research into the Navier-Stokes problem. According to OpenAI's research announcement, multiple agent groups using in-house models tackled the problem with different approaches. Among them, the group that reached a solution reportedly had about 10,000 agents running simultaneously. During the research process, Codex was used to compile useful findings from each group and share them across groups.

External service integration expanded through plugins

Corey Ching, who leads Developer Experience at OpenAI, gave a talk titled "ChatGPT Platform with Plugins," explaining the mechanism for finding plugins, the operation screens within ChatGPT, and the flow from development to publication. Behind this are developers' desires to have users discover their own plugins and to make the application and review process smoother.

What helps with discovery is a feature in which ChatGPT suggests necessary plugins according to the content of the conversation. On a slide, in response to a consultation about wanting to turn an existing website into a web app, the Supabase and Vercel plugins were presented along with explanations of their uses. An install button was also displayed on the conversation screen, showing an example in which the user can proceed to install on the spot.

The mechanism for embedding plugin operation screens into ChatGPT is Plugin Extensions. Users can open apps from the sidebar, display operation screens next to the conversation, and handle corresponding files in a dedicated viewer.

In a demo of "Rosalind Workbench," a work environment for life science research, Ching combined analysis through conversation with a viewer that displays the three-dimensional structure of molecules. A location selected in the viewer is also communicated to ChatGPT, allowing it to analyze or explain that part.

Demo of Rosalind Workbench. A molecular structure viewer is displayed next to the conversation.

In the design tool Magic Path, he requested a landing page for a baseball event by voice and displayed the generated result in the side panel. He then manually adjusted fonts and colors on the screen.

He also introduced examples of using Canva and Shopify to proceed with design production and store data confirmation within ChatGPT.[6]

MCP, which serves as the foundation for such integration, is a protocol for AI applications to interact with external data and tools. In the official documentation, the feature for embedding operation screens is presented as OpenAI MCP Extensions, which extend MCP and MCP Apps.

Ching also took up MCP Events, which trigger agents based on updates to external services. The official documentation explains a mechanism in which users specify what to monitor and how to respond to updates, and ChatGPT performs work upon receiving notifications such as comments on documents.

As a way to reduce the effort of building and operating MCP servers, Ching also took up the use of Sites. Sites is a service for creating and publishing websites and apps, and it can also host MCP servers. On a slide, he showed an example of creating a shared TODO list that agents can view and update, and requesting everything from deployment to Sites to packaging as a Codex plugin.

Example of requesting everything from creating a shared TODO list to deployment to Sites

Ching also mentioned that the number of app and plugin applications is increasing, and explained the application and review procedures.

Slide showing the increase in app and plugin applications in 2026

To publish a created plugin, upload the package in ZIP format, respond to issues pointed out by automated checks, and proceed to review. Under the current application procedure, after approval in review, developers can choose the timing of publication.

Bringing ChatGPT accounts and usage quotas to external apps

Ching also took up Sign in with ChatGPT, which allows users to sign in to external apps with their ChatGPT account. In addition to sign-in, in supported apps, users can use the usage quota of their ChatGPT plan for AI processing[7]. In the talk, he explained that this mechanism reduces the burden of inference costs on app developers and makes it easier for users to try external apps.

After this, sessions continued on designing, implementing, and using agents. There were also opportunities for participants to interact with the OpenAI team and other developers, making it a place to exchange knowledge about building things with AI.