# Live from OpenAI DevDay 2026: Keynote

## Executive summary

OpenAI unveiled major advancements across AI agents, model performance, and developer tooling. Key launches include 'Dots,' always-on AI agents designed to handle complex, ambitious tasks like tracing dependencies and writing code. The platform is enhanced by 'ChatGPT Space,' a collaborative environment for teams and agents. Technically, the focus was on speed and capability, introducing GPT-6.1 for cost-effective intelligence, 'UltraFast' for rapid token generation (up to 300 tokens/sec), and the 'Decisions API' for fast, focused model responses. Developer access is expanded with 'Codex in the Cloud,' 'Private Tensors (Z-D.R.)' for enhanced privacy, and 'API Canvas' to facilitate multi-agent workflows.

## Key takeaways

- AI Agents (Dots): Dots are always-on agents designed to handle ambitious, multi-step tasks, such as monitoring bug reports, testing fixes in draft PRCs, or tracing dependencies to migrate an app off a legacy API. They function as a 'second brain' and can operate across the cloud.
- Collaborative AI (ChatGPT Space): A new dedicated space for teams and agents to work together, allowing for interactive charts, live prototypes, and delegating tasks to multiple Dots simultaneously.
- Model Performance & Speed: Launch of GPT-6.1, offering a balance of cost and capability. 'UltraFast' was introduced across the API, ChatGPT, and Codex, providing significantly increased speed (up to 300 tokens/sec). The 'Decisions API' allows models to respond quickly by focusing on a pre-defined set of options.
- Developer Platform Enhancements: Codex is now available in the cloud, allowing tasks to start on a phone and continue on a desktop. New features include 'Private Tensors (Z-D.R.)' for frontier model safety without storing content on OpenAI's servers, setting a new standard for privacy.
- Multimodality and Computer Use: The Astra model demonstrates advanced capabilities, including interpreting sketches, building apps, and interacting with the operating system (computer use) in real-time, showcasing the integration of voice, vision, and code.

## Technical details

- AI Agents: Dots are designed to take on delegated responsibilities, acting as virtual teammates. They can follow user access, operate in the cloud, and write/test code, moving beyond simple tasks like booking tickets.
- Model APIs: The 'Decisions API' allows models to respond rapidly by constraining the model's output to a pre-defined set of options. 'UltraFast' provides significantly increased speed for API calls.
- Development Workflow: Codex is now available in the cloud, enabling seamless task continuation across devices. The 'API Canvas' and 'Multi-agent controls' facilitate complex, multi-agent workflows.
- Security and Privacy: Introduction of 'Private Tensors (Z-D.R.)' to deliver frontier model safety while ensuring user content is not stored on OpenAI's servers.

## Practical implications

- Build engineers can leverage Dots to automate complex, multi-step development tasks, such as dependency tracing and running PRs.
- The cloud availability of Codex allows for continuous development workflows, enabling tasks to be started on mobile and completed on desktop.
- The combination of UltraFast and the Decisions API provides developers with unprecedented control over model speed and output structure, critical for real-time applications.
- The integration of multi-agent controls and the API Canvas allows for building sophisticated, collaborative systems that mimic human team efforts.

## Topics

AI Agents, Large Language Models (LLMs), Software Development, Cloud Computing, Multimodality, Dots, ChatGPT Space, GPT-6.1, UltraFast, Decisions API, Codex in the Cloud, Private Tensors (Z-D.R.), Astra

Source: https://www.youtube.com/watch?v=Fls_onRviPM