Language: 한국어판
OpenAI’s DevDay 2026 was not just a model launch. Across more than 20 announcements, the company laid out a broader product architecture: AI that persists across sessions, runs in the cloud, uses software, and can take responsibility for ongoing work.
The clearest example is Dots. OpenAI describes Dots as always-on agents powered by GPT-6 Astra, each with its own cloud computer and access to connected apps. A Dot can keep working on projects in the background, learn from feedback, and surface completed work for review. OpenAI says the product can connect to more than 4,000 apps through its plugin ecosystem and is beginning to roll out to Pro, Business Premium, and eligible Enterprise users.
That matters because it changes the unit of interaction. The familiar chatbot waits for a prompt, produces a response, and largely ends there. A persistent agent is designed around a longer-lived goal: monitor something, continue a project, use tools, recover context, and return when work is ready or a decision is needed.
The same pattern is appearing in the developer stack
OpenAI’s new Agents API makes this architecture available to developers. The official documentation describes durable sessions in which OpenAI can manage orchestration, context compaction and recovery while an agent works inside a sandbox, edits files, connects to MCP servers, or uses a hosted browser. Computer use adds the ability to interact with software through its interface.
Codex is moving in the same direction. DevDay added cloud execution, reusable development environments, multi-agent workflows and code review that can continue away from a developer’s laptop. Plugin extensions, meanwhile, let third-party software appear inside ChatGPT surfaces rather than living only behind a conventional app boundary.
The model layer is being optimized for this shift too. OpenAI says GPT-6.1 Sol approaches GPT-6 Astra on agentic coding, computer use and professional work while charging one-fifth of Astra’s standard input and output token prices. Its published standard API pricing is $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. Lower inference cost matters more when software is expected to run repeatedly and in the background rather than answer one isolated question.
What DevDay does not prove
Product architecture is not the same thing as reliable autonomy. The performance comparisons above are OpenAI’s own evaluations, and OpenAI explicitly notes that research or API evaluations can differ from production ChatGPT because system prompts, tools and reasoning settings differ. The computer-use documentation also requires applications to handle website approvals and sign-in, review browser activity and verify results.
So the important conclusion is narrower than “agents have arrived.” DevDay shows where OpenAI is placing its product bets: persistent identity, durable sessions, cloud execution, software access and a growing ecosystem around agents. Whether those systems can be trusted with long-running real-world responsibilities at scale remains an empirical question.
For developers and organizations, that changes the planning question. The next AI project may not begin with “Which model should answer this prompt?” It may begin with “Which responsibility can an agent hold over time, what tools may it use, and where must a human remain in control?”