OpenAI DevDay 2026: Everything Announced
OpenAI DevDay 2026 introduced a substantial expansion of the company’s platform, spanning persistent AI agents, collaborative workspaces, new models, faster inference, cloud-based Codex, new agent infrastructure, privacy controls, and new ways for developers to distribute and monetize applications. More importantly, the announcements outlined a broader direction for OpenAI. The company is moving beyond AI systems that simply respond to individual prompts and toward agents that remain active, maintain context, operate software, and perform work continuously on behalf of users.
You can see the live replay here
Here is everything OpenAI announced at DevDay 2026.
Dots: Always-On AI Agents
The centerpiece of DevDay was Dots, a new type of persistent AI agent designed to work continuously on a user's behalf. Instead of waiting for individual prompts, a Dot can be given an ongoing responsibility and continue working on it over time. OpenAI demonstrated Dots monitoring incoming information, maintaining presentations as new data arrived, coordinating calendars, working across Slack, investigating bugs, preparing pull requests, testing software changes, managing migrations away from legacy APIs, and keeping projects on track.
A Dot gets its own computer in the cloud and its own browser, and it can write and test code when necessary. It follows the user's existing access permissions and can use the same plugins that the user has connected to ChatGPT. According to OpenAI, Dots can work with more than 4,000 applications through ChatGPT. They are powered by Astra, and users can configure boundaries around application and computer use as well as instructions governing the actions a Dot can take.
OpenAI also wants Dots to operate outside ChatGPT. The company said users will eventually be able to text their Dot, receive proactive updates by text, and call it directly. OpenAI is initially allowing users to work with one Dot, but the longer-term plan is to allow a user to work with an entire team of them.
The key idea is that a Dot is not simply another chat interface. OpenAI presented it as a persistent delegate that knows what it has been assigned, keeps working without requiring continuous prompting, and asks the user for input when a decision or approval is required.
ChatGPT Space and Team Collaboration
While Dots are centered around an individual, ChatGPT Space is designed around collaboration. Space is a shared environment where people and their agents can work together on pages and files. Users can write plans, conduct research, generate images, analyze data, create interactive charts and sheets, and build live prototypes. Teammates and Dots can then be brought directly into that work.
Agents can be tagged in pages and comments much like human collaborators. A Dot could, for example, be instructed to monitor a particular Slack channel and update a page every day with its findings. OpenAI also demonstrated a Dot modifying an interactive visualization, finding information from a user's Slack messages, and incorporating feedback from other people into a shared document.
Slack itself is becoming an important environment for Dots. Users can forward a thread to their Dot and delegate the work that comes out of it. Multiple people can also bring their respective Dots into a group conversation, where the agents perform tasks while keeping everyone informed. OpenAI showed an internal example where an engineer's Dot detected a reported bug, investigated it, and opened a pull request with a fix. The company said its engineers are already using Dots to fix dozens of bugs per day through workflows like this.
OpenAI is also bringing presentations into Space. The upcoming presentation system is being designed so that slides can be easily read and modified by both humans and agents, allowing a team and its Dots to collaborate directly on a presentation.
Both Dots and Space are available to ChatGPT Pro, Business Premium, and Enterprise users. OpenAI said a user's Dot is included with the subscription and conversations with it do not consume the user's normal usage allowance.
Specialist Dots for Enterprise
Enterprise customers are also getting Specialist Dots, which extend the concept from personal agents to organization-wide virtual teammates. A company can configure these agents for specialized functions such as accounting, marketing, or legal work, provide them with shared goals and context, and review their output. Feedback can then be incorporated across the organization rather than remaining specific to one user.
OpenAI is also working with Microsoft to integrate Specialist Dots into Agent 365. This will allow businesses to manage their Dots through Microsoft's existing enterprise tooling.
GPT-6.1 Sol and Ultra-Fast Inference
On the model side, OpenAI announced GPT-6.1 Sol, which it positioned as a high-capability model intended for frequent everyday use. OpenAI said Sol provides intelligence close to Astra at approximately one-fifth of the price and is smarter than Astra in some areas. Cached input also receives a 95 percent discount compared with standard input, which is particularly relevant for agents that repeatedly reuse large amounts of context while working through long-running tasks.
OpenAI described GPT-6.1 Sol as a model that developers can use as their everyday workhorse. Rather than always paying for the company's highest-end model, developers can use Sol for workloads that require strong reasoning and coding capabilities at a substantially lower cost.
OpenAI also introduced Ultra-Fast, a new inference tier available across the API, ChatGPT, and Codex. Standard inference serves as the baseline, while Fast provides twice the speed at twice the price. Ultra-Fast reaches eight times the standard speed at six times the standard price, with OpenAI claiming approximately 300 tokens per second. Ultra-Fast is initially available with Astra, with GPT-6.1 Sol support planned later.
These changes also come with new ChatGPT subscription options. Pro 500 provides OpenAI's highest usage limits and access to Ultra-Fast in ChatGPT and Codex. According to the keynote, it includes 25 times the usage of Plus and can be used across supported partner applications through Sign in with ChatGPT. OpenAI is also reopening Pro 200, which continues to provide access to the company's frontier models.
Decisions API
OpenAI previewed a different approach to low-latency inference with the Decisions API. Instead of asking a model to generate an unrestricted response, developers provide a predefined set of possible choices and the model selects among them. By constraining the problem this way, OpenAI says the model can respond in a fraction of a second.
The Decisions API uses OpenAI's Luna model and retains capabilities such as image understanding, broad language support, and safety protections. OpenAI suggested applications such as routing requests, classifying images, selecting which agent should handle a task, and deciding what an agent should do next. The company also highlighted robotics as a potential use case, where a system could process visual input and choose an action with very low latency.
OpenAI's AI Research Intern
OpenAI also announced that it had reached a milestone it predicted the previous year: an AI research intern. The company describes this as a system capable of independently completing a clearly defined research task that would otherwise require substantial time and effort from a skilled researcher.
OpenAI said that in January its models performed well on research tasks taking less than 15 minutes but usually failed on work that took a day or longer. By July, the models could complete more than one-third of day-long internal research tasks without human intervention. OpenAI also described using its own models in continuous research loops to improve computer-use systems, including optimization work that resulted in more than a 2× latency improvement.
The company said models such as Astra have contributed to research involving mathematics, antibiotics for drug-resistant infections, ancient languages and history, renewable energy, energy efficiency, robotics, and manufacturing. OpenAI said Astra-related models have also helped solve more than 100 mathematical problems that had remained open for decades.
Codex Goes Fully to the Cloud
Codex received several major updates at DevDay. The first is that Codex is now fully available in the cloud. A developer can start a task on a phone, continue from a browser or desktop, close their laptop, and allow the agent to keep working remotely. OpenAI demonstrated the system by asking Codex to rewrite an application's entire backend in Rust and then leaving the task running in the cloud.
The cloud version supports the same broader tool environment as Codex running locally, including plugins and computer use. This makes long-running coding tasks independent of the developer's local machine and allows work to move between devices without losing the running agent or its context.
The infrastructure behind Codex is the Codex harness, which also powers Work and Dots. OpenAI announced that this harness is open source. The company said it has optimized the system to produce accurate results quickly while minimizing unnecessary token usage, and developers can now inspect it and use it as a foundation for their own agent systems.
Codex Security Cloud and the Agents API
OpenAI also launched Codex Security Cloud, a service designed to continuously inspect software for vulnerabilities. It can identify vulnerabilities, automatically deduplicate findings, perform scheduled scans, and prepare verified fixes for developers to review. The goal is to provide security teams with agents that continuously inspect and harden their infrastructure rather than relying entirely on periodic manual reviews.
The same broader agent infrastructure is being exposed through the new Agents API, which launched in public beta. It includes the agent harness, hosting, memory, multi-agent controls, computer use, and other components used by OpenAI's own products. OpenAI described it as the same underlying technology that powers Codex, Dots, and its other agents.
Computer use is particularly important here because it allows an agent to operate software rather than interacting only through APIs. OpenAI demonstrated an agent opening a browser, navigating through a website, clicking through pages, and testing a user flow. This allows developers to build agents that can interact with software in much the same way a person would.
OpenAI is also expanding its partnership with AWS. Developers can build Bedrock managed agents powered by OpenAI, allowing OpenAI agents to operate alongside applications and data already running within AWS. AWS customers can also access OpenAI frontier models, Codex, and ChatGPT Work.
OpenAI Private Intelligence
For organizations working with sensitive information, OpenAI previewed OpenAI Private Intelligence. The announcement includes Zero Data Retention with private safety processing, which allows OpenAI's frontier-model safety systems to operate without storing customer content on OpenAI's servers.
OpenAI is also introducing private inference, which is intended to provide privacy protections while inference itself is taking place. Together, OpenAI positioned these technologies as a way for organizations to apply frontier models to sensitive workloads while maintaining stronger controls over their data.
Responses API Performance Improvements
The underlying API infrastructure also received significant performance improvements. OpenAI said Responses API usage grew 100× during the previous year while reliability remained above 99 percent. During the same period, the company reduced time to first token by 45 percent and made tool calls and workflows more than 30 percent faster.
These improvements are increasingly important as API usage shifts from individual model requests toward agents that perform long sequences of model calls, tool calls, and external actions. Latency and reliability across an entire workflow become just as important as the speed of an individual model response.
A Major Codex CLI Update
The Codex CLI has been substantially refreshed and now includes bidirectional audio directly inside the terminal, powered by GPT Live. This is not simply voice dictation. Developers can have an ongoing voice conversation with Codex while working from the command line and receive responses in real time.
OpenAI combined the new CLI with Ultra-Fast during the keynote to demonstrate creating and modifying an application in seconds. The demo showed how the faster inference tier changes the interaction model when coding with an agent, since developers can request a modification and see it implemented almost immediately.
Codex is also gaining deeper computer-use capabilities through App Shots. An App Shot provides Codex with context from another application and allows it to interact with the computer. In the demonstration, Codex was asked to audit an application across different screen sizes. It navigated the relevant applications and simulators, inspected the interface, and captured screenshots as it worked.
Together with cloud execution, these features move Codex beyond generating and editing source code. The agent can increasingly build software, run it, interact with it, test it, and continue working without requiring the developer to supervise every step.
Multimodal Development
OpenAI also demonstrated how its different models can be combined within a single application. A programmable robot called Lavender used Astra for vision, GPT Image 2.5 for image generation, and GPT Live 1 for real-time voice interaction. OpenAI also discussed using the Decisions API to allow robots to make fast decisions based on visual input.
The demo was intended to show that OpenAI's text, vision, audio, image generation, computer-use, and agent capabilities can increasingly be treated as interoperable building blocks. Developers can combine them according to the requirements of an application rather than treating each modality as an isolated product.
Bring Your ChatGPT Subscription to Other Apps
OpenAI introduced a major change to how developers can fund AI usage in third-party applications. Users will be able to bring their existing ChatGPT subscription into supported third-party products.
This means a developer does not necessarily have to pay the token costs associated with getting an existing ChatGPT subscriber started on their application. Eligible AI usage can instead come from that user's ChatGPT plan. OpenAI launched the capability with 16 partners and said it intends to expand the program quickly.
Plugin Extensions
OpenAI is also substantially expanding what developers can build as a ChatGPT plugin. With Plugin Extensions, a plugin can become an application that runs natively inside ChatGPT and Codex rather than simply exposing tools that the model can call.
Developers can use Plugin Extensions to create editors, dashboards, workspaces, and other interactive interfaces. OpenAI demonstrated a Meetings application that displays upcoming calendar events inside ChatGPT and allows users to start taking notes directly in a Space. Figma demonstrated opening designs and making revisions from ChatGPT, while Adobe demonstrated Photoshop functionality operating directly within the interface.
Plugin Extensions can also be integrated into ChatGPT Sites. A site can connect to external applications and data while individual visitors authenticate with their own credentials and bring their own agent. This allows the same application to provide personalized functionality and access depending on the person using it.
Conversational Plugin Discovery
OpenAI is also changing how users discover third-party software. Plugins will remain searchable through the plugin library, but ChatGPT will now be able to surface an appropriate plugin during a conversation when it recognizes that the plugin could help with the user's request. The user can then connect it without leaving the flow of the conversation.
The plugin submission process is also being revised. Developers will be able to track the status of a review, see what needs to be corrected, request human review, and update individual plugin tools without restarting the entire submission process.
OpenAI Marketplace
Finally, OpenAI launched the OpenAI Marketplace, giving developers another distribution channel for products built around its ecosystem. The Marketplace launched with more than 30 partners, including CodeRabbit, Notion, and Vercel.
Enterprise customers can use part of their existing OpenAI financial commitment to purchase third-party products through the Marketplace. Companies with their own OpenAI commitments can also spend those commitments there. Through a partnership with Base10, OpenAI said customers will also be able to access open-source models through the Marketplace.
OpenAI also briefly announced an OpenClaw Enterprise Harness as part of its continued investment in open source. The keynote did not provide technical details because a separate DevDay session was scheduled to cover it.
OpenAI DevDay: The Bigger Picture
DevDay 2026 included a large number of individual launches, but they fit together into a fairly clear platform strategy. Dots provide persistent agents, while Space gives humans and those agents somewhere to collaborate. GPT-6.1 Sol, Astra, Ultra-Fast, and the Decisions API provide different combinations of intelligence, cost, and latency. Codex Cloud and the Agents API provide execution infrastructure, while computer use and plugins give agents access to existing software and services.
Private Intelligence addresses the problem of applying these systems to sensitive workloads. Plugin Extensions, ChatGPT subscription portability, conversational plugin discovery, and the OpenAI Marketplace address the other side of the ecosystem by giving developers ways to build products, reach users, and sell them.
The common thread is a shift away from AI as a model that simply responds to prompts. At DevDay 2026, OpenAI presented AI as persistent infrastructure. Agents can remain active, maintain context, operate software, collaborate with humans and other agents, and continue working after the user has moved on to something else. The individual announcements matter, but the larger story of DevDay is the platform OpenAI is assembling around them.
Comments