OpenAI DevDay 2026: Dots Agent, ChatGPT Space, GPT-6.1 Sol

OpenAI held DevDay on September 29, launching Dots, an always-on personal agent powered by GPT-6 Astra that proactively manages work and personal tasks and seeks approval only at key checkpoints. It also introduced ChatGPT Space, a human-agent workspace, and GPT-6.1 Sol at one-fifth of Astra's price.
Key insights
Dots: an always-on personal agent powered by GPT-6 Astra
OpenAI introduced Dots at DevDay 2026, a resident agent with its own cloud computer and browser that can write code and run tests. Dots proactively advances tasks and only seeks user approval at key checkpoints, managing both work and personal matters such as board deck edits, code deployments, cake orders, and scheduling conflicts. It connects to more than 4,000 apps already linked in ChatGPT, can later be reached by SMS and phone calls, and runs on GPT-6 Astra, which Sam Altman called OpenAI's best-aligned model. Each user starts with one Dot, with sets planned later, and users can restrict which apps it uses and write custom instructions for its actions.
ChatGPT Space: a workspace where humans and agents collaborate
ChatGPT Space is organized around pages where users can write plans, do research, and generate images and data, with pages and files stored in one space like a cloud drive. Users can @ their Dot in comments to assign work, and pages can carry standing instructions such as checking a Slack channel daily. Slash commands insert interactive charts, tables, and runnable prototypes; in the demo, a Dot named Dotty converted data into a filterable bar chart refreshed hourly and pulled onboarding data mentioned in a Slack DM into an FAQ. Presentation support is planned, with the format designed for agent readability.
Dots already change how OpenAI works internally
At OpenAI, Dots have their own identities in company Slack, and employees treat them as proxies, forwarding scattered requests and adding their Dots to group chats from the start. In one demonstration simulating a launch eve for a playlist app called Blossom Music, Dotty rescheduled a moved-up review meeting, reviewed overnight tester feedback, and built a redesigned homepage using Codex on the presenter's laptop, running it in an iPhone simulator before submitting a PR. Engineers' Dots pick up bugs posted in feedback channels and submit fix PRs; the company stated its Dots fix dozens of bugs daily, and many parts of the Dots product itself are written by Dots.
Enterprise Specialist Dots arrive with Microsoft Agent 365 integration
Dots and Space opened to ChatGPT Pro, Business Premium, and Enterprise users on launch day, with Dot conversations included in plans rather than charged against quota. Enterprise customers get a preview of Specialist Dots: company-configured virtual colleagues for accounting, marketing, or legal workloads, given goals and context by the company, with outputs reviewed and feedback shared company-wide. OpenAI is also working with Microsoft to connect Specialist Dots into Agent 365, Microsoft's enterprise agent management tool, letting companies manage them with existing Microsoft tooling.
GPT-6.1 Sol cuts cost to one-fifth of Astra; UltraFast hits 300 tokens per second
GPT-6.1 Sol offers capability close to GPT-6 Astra at one-fifth the price, with cached inputs 95% cheaper than standard input, targeting developers' daily workhorse needs; Altman said Sol is smarter than Astra in some respects. The new UltraFast speed tier delivers 300 tokens per second at six times the price for eight times the speed of standard, available in the API, ChatGPT, and Codex, and demoed building a DevDay-themed rocket faster than standard speed. A $500 Pro 500 subscription offers 25 times Plus quota and UltraFast access, while Pro 200 reopens with all frontier models. A Decisions API preview constrains the Luna model to picking from predefined options, cutting response times under one second.
OpenAI says models now help run its own research
Post-training research lead Tejal presented models improving OpenAI's own research pipeline: one model optimized the computer-use harness in a loop, producing changes that more than doubled latency improvements in production, and another improved monitoring and refusal training, leading Astra to industry-leading computer-use stress test results. Internal data showed research token consumption rising sharply after this summer; in January models handled tasks up to 15 minutes, while by July more than a third of day-long research tasks completed without human intervention. Astra-class models have helped resolve over 100 open mathematics problems decades old and contribute to antibiotic research against resistant infections, ancient languages, renewable energy, and industrial robotics.