OpenAI DevDay 2026: the launches and what changes for people building with AI

This article brings together highlights from the official announcement OpenAI sent to the community on the day of the event. The launch descriptions and prices below follow that material. The analysis of applications for founders, developers and companies is NewHack's.

The main launches at OpenAI DevDay 2026

• Dots: agents with ongoing responsibilities, context from connected apps and a computer in the cloud.
• ChatGPT Space: collaboration between teams, ChatGPT and Codex, bringing files and shared work together.
• GPT-6.1 Sol and Astra Ultrafast: developments focused on the relationship between capability, cost and speed.
• Plugin extensions: new ways to integrate products into the ChatGPT and Codex interfaces.
• Codex: reusable cloud environments, an evolving terminal experience, code review and security tools.
• Agents API and Decisions API: executing tasks in software and making structured decisions, with the Decisions API announced in preview.
• Private Intelligence, Pro 500 and shareable profiles: developments in data protection, usage limits and sharing creations.

Dots: agents that follow through on a responsibility

Dots were introduced as agents capable of taking on responsibilities defined by the user, working across conversations and finding context in connected apps. They can track work in progress and alert the person when something needs attention.

According to the announcement, dots use Astra, have their own cloud computer and can draw on Codex and ChatGPT Work to research, analyze data, prepare documents and develop software. Setup starts on a computer, through the ChatGPT app or browser, with continued use on mobile after that step.

One of the examples OpenAI presented is producing content from an interview. When a transcript arrives, the dot can select excerpts for short videos, prepare notes and suggest posts, incorporating adjustments into subsequent materials.

For a startup, one possible application would be following customer discovery interviews. The agent could organize evidence, highlight unanswered questions and prepare a synthesis for the product meeting. The team would need to define which sources it can consult, when it should request review and what constitutes a useful deliverable.

ChatGPT Space: shared context for the team

ChatGPT Space brings together collaboration between people, ChatGPT and Codex, alongside uploaded files, creations and shared work. The aim is to make materials easier to find and work easier to resume.

In our view, this addresses a common problem in companies adopting AI: learning becomes scattered across individual conversations. A shared space can help the team find the relevant version of a document and understand what has already been produced. Making it work well still requires organizing ownership, permissions and references.

GPT-6.1 Sol and Astra Ultrafast: cost and speed become part of the decision

In the section on models, the announcement presents GPT-6.1 Sol as an advance for agentic coding, computer use and professional work, with intelligence close to Astra's at one-fifth of the price. This is the comparison published by OpenAI, rather than an independent assessment by NewHack.

The standard API prices stated for GPT-6.1 Sol in the announcement are:

Tokens are units used in the model's processing and billing. The cached input price represents a 95% discount on standard input for context that can be reused in this way. That does not amount to a 95% reduction in the entire bill: output, tools and other workflow stages also affect the cost.

The material also announces Astra Ultrafast as six times faster than standard Astra. The actual gain in an application depends on the work performed, including time spent using tools and external systems. Availability and access conditions should be checked in the product and current documentation.

For founders, the implication is economic: a workflow requiring several interactions may become viable with a different combination of model, context and speed. The relevant test is how much it costs to complete a task at an acceptable quality level, including errors and human review.

Plugins gain more space within ChatGPT and Codex

Extensions let parts of a product become part of the ChatGPT and Codex experience. The announced possibilities include a space in the sidebar, an interactive panel beside the conversation and a viewer for a file type. OpenAI also reported improvements in creating, submitting and discovering plugins.

This opens up a potential distribution path for software companies: offering a useful action in the environment where the user is already working. An analytics product, for example, can present results in a panel that allows data exploration without losing the context of the conversation.

The opportunity needs to be tested with users. Being available as a plugin does not guarantee adoption. The value lies in solving a specific need in fewer steps while keeping permissions and the actions to be performed clear.

Codex advances from creating software to reviewing and securing it

The new Codex features cover several stages of development. The announcement highlights:

• Codex in the cloud: reusable environments, progress tracking and the ability to steer and resume work from a phone or another computer.
• Codex CLI: voice to start and steer tasks, agent tracking in /agents and worktrees, which separate code workspaces, within a renewed terminal experience.
• Codex Review: reviewing GitHub pull requests and GitLab merge requests in the desktop app, with questions about changes and configuration of automatic reviews.
• Codex Security Cloud: scans of entire GitHub repositories, on demand or on a schedule, and continuous checks of new commits.

For small teams, it is worth experimenting with a complete workflow: start from a user need, implement a change with a defined scope, run the checks and review the result. The measure of success is a change that works and can be maintained by the team.

This topic already appears in NewHack's work with internal tools built with AI and in the experience shared by Rodrigo Terron as a Codex Ambassador. Context, acceptance criteria and checking the result are part of that process.

Agents API and Decisions API: executing tasks and choosing the next step

The Agents API now supports computer use, according to the announcement. This allows the creation of agents that interact with software to perform tasks. The announcement also highlights Codex capabilities for multiple agents, tool search and calling, and context compaction, with infrastructure operated by OpenAI.

The Decisions API, announced in preview, receives text or images and a set of possible answers to return structured results. Cited uses include classifying content, routing requests and choosing an agent's next action.

One possible application would be routing a support request to a predefined category and then triggering a specific workflow. In NewHack's analysis, that separation helps test each stage: was the classification correct? Did the agent perform the expected action? Should the case have been routed to a person?

Private Intelligence and the announcement's other developments

Under Private Intelligence, OpenAI announced Zero Data Retention with Private Safety Processing, described as a way to perform automated, offline safety analysis without retaining customer content. It also introduced the preview of Private Inference, combining confidential computing with verifiable controls.

For a company evaluating AI with internal data, these options should be part of the architecture analysis and project requirements. Eligibility, scope and the conditions applicable to the chosen configuration need to be verified before treating the announcement as a guarantee for any use.

The new Pro 500 plan was presented with the highest usage limits in the announced offering and access to Astra Ultrafast in Codex and ChatGPT Work. Shareable profiles bring together creations such as Sites, pages and skills to make them easier to present to other people.

NewHack's interpretation: test a complete deliverable

The announcements bring continuity, collaboration, software building and product integration closer together. For entrepreneurs, the practical question is which part of the work can improve now and how to measure that change.

• Choose a recurring routine, such as preparing an analysis of interviews, reviewing a code change or classifying requests.
• Record how it works today: time spent, rework, cost and quality of the result.
• Define the available data, permitted actions and moments for human review.
• Test with real examples and compare the complete result, including failures and exceptions.
• Share what you learn with the team before expanding usage.

For companies, this path can start with a well-defined AI use case or with a hackathon connected to a business challenge, provided someone is responsible for evaluating the prototype and leading the next stage.

It is also an opportunity for the ecosystem: developers, founders and companies learn more when they can compare attempts, discuss limits and build together. AI House Brasil is one of NewHack's initiatives focused on that connection. Follow our community to turn the launches into conversations, experiments and projects that can be evaluated in practice.

Sources and context for this coverage

Main source: the official OpenAI DevDay 2026 announcement sent to the community and received by NewHack on September 29, 2026. Names, features and prices were presented according to that material; use examples and interpretations were identified as NewHack's analysis. Decisions API and Private Inference were announced in preview. This coverage records the announcement, rather than an independent test of the products.

Additional official references: OpenAI DevDay 2026 announcement and OpenAI documentation on agents. Check the current conditions before implementing or purchasing access to a feature.