Product news: OpenAI DevDay 2026 expands agents, developer tools, and shared AI workspaces

OpenAI held DevDay 2026 in San Francisco on September 29, announcing updates across models, agents, developer tools, and collaborative work. For education and research, the significant connection is between these layers: a model can handle a task, an agent can continue working with tools, and a shared workspace can expose the resulting artifacts to colleagues. The announcements create possibilities for new workflows, without establishing that those workflows improve learning.

GPT-6.1 Sol was a central model release. OpenAI describes improvements in coding, computer use, document understanding, and scientific workflows, with standard input and output token prices at one-fifth of GPT-6 Astra's. Its release page includes evaluations of data analysis, simulation, and other scientific tasks. Those results help identify candidate uses, but they remain bounded evaluations. A research group should test its own methods, source requirements, and failure cases before treating the model as a dependable assistant.

Dots adds a different kind of capability: ongoing work through an agent with its own cloud computer and connected applications. OpenAI's examples include a scientist receiving updated analyses and figures when new evidence arrives. This is a product scenario, not an independently evaluated research outcome. For a lab, the practical questions are which data the agent may read, which operations it may perform, and which changes require human approval. An automated refresh must not silently replace an accepted analysis.

DevDay also extended the Agents API with computer use. The underlying API had already entered public beta on September 10; September 29 should not be reported as its original launch. It supports managed execution, tools, context handling, and subagents, while the new interface capability broadens the software an agent can operate. Educational developers could explore bounded administrative or research tasks, but should preserve logs and test both successful actions and refusals outside the permitted scope.

Other announcements included plugin extensions, ChatGPT Space, and Pages for shared work. Collaborative slides were announced for the following weeks, and Decisions API was described as a limited preview on the event date. These distinctions matter when planning a course or institutional pilot. An announced feature is not necessarily available to every student, device, region, or subscription, and a shared document does not itself establish clear authorship or consent.

Anthropic had released Claude Sonnet 5.5 the previous day, positioning it as a faster, lower-cost complement to Opus 5.5 for bounded work, coding fixes, and documents. That contemporary alternative keeps the comparison focused on task fit rather than ecosystem allegiance. Gemini's previously released Guided Learning provides another reference: stepwise explanations and study aids aim to involve learners in working through a topic. It was an existing educational capability, not a DevDay announcement.

For Hong Kong educational teams, a useful first pilot would connect one approved source collection to one carefully specified task, then require a human to inspect the output and its supporting evidence. Researchers might compare an agent's report against a manually checked analysis; teachers might assess whether students can justify the choices made in an AI-assisted artifact. The event broadens what can be built and delegated. Its educational significance will depend on accessible participation, accountable permissions, and demonstrable understanding after the automated work is complete.