OpenAI DevDay Expands From Models to the Enterprise Agent Stack

Key Highlights

• OpenAI is moving beyond models and APIs into the runtime, interface, and infrastructure enterprises use to build and operate agents.
• Dots introduce persistent agents that maintain context, work across applications, and take on ongoing responsibilities rather than responding to individual prompts.
• The expanded Agents API gives developers managed computer use, tool selection, multi-agent orchestration, and context management.
• The Decisions API provides constrained outputs for routing, classification, and workflow decisions that do not require open-ended text generation.
• Faster development comes with a larger architectural question: how easily can enterprises move models, agent state, evaluations, policies, and workflows if OpenAI controls more of the stack?

The News

OpenAI used DevDay 2026 to announce more than 20 additions across its models, APIs, Codex, ChatGPT, and enterprise products. The most consequential announcements for enterprise developers include Dots, GPT-6.1 Sol, an expanded Agents API, the Decisions API, Private Intelligence, Bedrock Managed Agents powered by OpenAI, and new collaborative work surfaces in ChatGPT. Together, the announcements extend OpenAI beyond model access into agent orchestration, development environments, workplace interfaces, identity, and distribution. More details are available in the OpenAI DevDay 2026 recap.

Analyst Take

The most important DevDay announcement was not a single product. It was the shape of the portfolio.

OpenAI is assembling more of the environment in which enterprise agents are built, deployed, governed, and used. The company now offers the models, agent harness, hosted runtime, computer-use capabilities, coding environment, collaborative workspace, plugin ecosystem, and persistent agent interface. This could make agent development substantially easier. It also gives OpenAI influence over more architectural decisions that previously belonged to application developers, cloud providers, or enterprise software vendors.

Dots represent the clearest change in how OpenAI wants people to use AI. A chatbot waits for a question. A dot maintains context, monitors connected systems, works in the background, and takes on responsibilities over time. OpenAI says Dots have a cloud computer, can connect to more than 4,000 applications, and can carry context across ChatGPT, Slack, and Microsoft Teams.

That persistence is useful, but it changes the risk model. An agent that operates continuously needs a durable identity, carefully scoped permissions, limits on what it can read and change, and a reliable record of its actions. Enterprises also need to know what the agent remembers, how that memory changes, and what happens to its accumulated context when an employee changes roles or leaves the organization.

OpenAI has started addressing those questions. Dots use separate cloud computers, support custom rules and approvals, and restrict proactive background research to read-only tools. OpenAI is also previewing specialist dots with dedicated identities and credentials. Those are the right architectural elements. The harder work will be proving that permissions remain consistent as a dot moves between ChatGPT, Slack, Teams, browsers, plugins, and systems of record.

For developers, the Agents API removes a substantial amount of plumbing. OpenAI is packaging tool search, tool calling, multi-agent delegation, computer use, context management, and the Codex agent harness behind a managed service. Developers can spend less time assembling orchestration components and maintaining agent infrastructure. The trade-off is that more application behavior now depends on OpenAI’s abstractions, execution environment, and pricing.

Agent portability is more complicated than model portability. An enterprise may be able to replace GPT-6.1 Sol with another model, but that does not mean it can easily move the agent. Tool definitions, context-compaction behavior, memory, subagent coordination, evaluations, approval logic, and audit records may all be tied to the original runtime. According to the 1H 2026 HyperFRAME Research Lens, 79% of organizations plan to use multiple foundation models. Enterprises should determine whether the Agents API helps them coordinate those models or quietly makes OpenAI the permanent center of the workflow.

The Decisions API may be one of the more practical announcements. It focuses a model on user-defined questions with a finite set of possible answers, returning outputs that applications can use to classify content, route work, or choose an agent’s next action. Enterprises have used probabilistic machine learning classifiers for years. The difference is the possibility of using one flexible model across many constrained decisions rather than developing and maintaining a separate classifier for every task.

That could reduce token generation, latency, retries, and output-validation work. It may also make agent behavior easier to evaluate because the expected output is bounded. But a constrained answer is not automatically a correct one. Developers still need confidence scores, fallback rules, testing against production data, and a way to route uncertain or high-impact decisions to a person or a more capable model.

GPT-6.1 Sol addresses a related economic problem. OpenAI says it approaches GPT-6 Astra’s performance on several coding, computer-use, and professional-work evaluations at one-fifth of Astra’s standard token prices. Cached input costs $0.10 per million tokens, which could be meaningful for agents that repeatedly reuse large amounts of context. These are OpenAI’s evaluation results, however. Enterprises should compare cost per successfully completed workflow, not token prices or benchmark scores in isolation.

Private Intelligence is also more substantive than a generic privacy promise. Zero Data Retention with Private Safety Processing allows automated safety review without OpenAI retaining customer prompts or responses. Selected encrypted records remain in customer-controlled cloud storage, and a hardware-attested environment performs the review without human access to the underlying content. This gives customers more control, although it also creates operational responsibilities around storage, encryption keys, permissions, regional configuration, and the required retention period.

The overall direction is clear. OpenAI wants to make building agents easier by owning more of the difficult parts. Enterprises now have to decide how much of that convenience they are willing to exchange for dependence on a single provider.

What Was Announced

Persistent agents. Dots are always-on agents powered by GPT-6 Astra. Each dot receives its own cloud computer and can connect to applications through OpenAI’s plugin ecosystem. Users can interact with a dot through ChatGPT, Slack, Teams, and voice. OpenAI says Dots can carry context between those environments and work on multiple projects without requiring the user to direct every step.

OpenAI has added an Activity View, custom action rules, automatic action review, and approval requirements for sensitive tasks. Proactive research uses read-only access when the user is not actively working with the dot. The company is also piloting specialist dots with separate organizational identities, credentials, and access to systems of record.

Managed agent infrastructure. The Agents API brings the harness behind Codex into customer applications. It includes tool search, programmatic tool calling, MCP support, custom functions, built-in tools, and multi-agent execution. Subagents can work in parallel with separate contexts while a primary agent coordinates the results. OpenAI operates the harness, while developers pay for the models and tools used by the agent.

Computer use extends that runtime from API calls into software interfaces. This expands what developers can automate, but it also increases the importance of sandboxing, permissions, action review, and replayable audit records. An agent that can operate a browser or desktop application can affect systems that were never designed for autonomous access.

Constrained machine decisions. The Decisions API accepts text or images and selects among a finite set of developer-defined answers. Likely uses include content classification, request routing, verification, policy checks, and workflow branching. The API is initially available in limited preview.

Model economics and speed. GPT-6.1 Sol is available through the API, ChatGPT Work, and Codex. OpenAI lists standard API pricing of $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. The new Ultrafast tier is designed for latency-sensitive work, with OpenAI claiming up to eight times faster token generation in Codex and up to six times faster generation through the API.

Private execution and cloud choice. OpenAI introduced Private Intelligence, including Zero Data Retention with Private Safety Processing and a preview of Private Inference. It also worked with AWS on Bedrock Managed Agents powered by OpenAI. AWS says these agents will use OpenAI capabilities while running within AWS and integrating with AWS resources. This gives AWS customers another route to OpenAI models and agent infrastructure, but enterprises should still examine where agent state, telemetry, evaluations, and support responsibilities reside.

Developer and workplace surfaces. Codex can now run in reusable cloud environments and continue working when the developer closes a laptop. OpenAI also added cloud code review, repository scanning through Codex Security Cloud, and improved multi-agent management in the command-line interface.

ChatGPT Space, Pages, collaborative slides, shared tasks, and integrations with Slack and Teams extend OpenAI into daily team workflows. Plugin extensions give developers persistent space in the ChatGPT sidebar and support interactive interfaces rather than conversational responses alone. These additions make ChatGPT both a development platform and a distribution channel.

Looking Ahead

OpenAI is competing for more than model usage. It is competing to become the environment where enterprise work is coordinated.

That puts the company into a complicated position with its partners. Microsoft owns workplace identity, productivity applications, and a large enterprise distribution channel. Salesforce owns valuable customer data, workflows, and business logic. AWS owns cloud infrastructure and many of the systems agents will need to access. OpenAI increasingly wants to own the agent runtime and the interface through which employees reach those systems.

The more interesting question is who owns the screen and the agent state behind it. A Salesforce record or AWS resource may still contain the authoritative data, but an employee could increasingly reach it through a dot, ChatGPT Space, Slack, or Teams. The system of record remains important. It may no longer determine how the user experiences or acts on its data.

Developers will benefit from a more complete agent platform. A managed harness, tool discovery, computer use, constrained decision APIs, and reusable cloud environments can shorten the path from prototype to working application. But a shorter path to deployment does not necessarily create a manageable production system. Enterprises still need to evaluate answer quality, tool selection, permissions, retries, loops, handoffs, cost, and business outcomes against real production traffic.

They should also test portability before it becomes urgent. Can a team change the model without rebuilding the agent? Can it export memory, workflow state, evaluation results, audit records, and policy configurations? Can the same controls work if the agent runs through AWS, Microsoft, or another provider? Access to multiple models is useful. The ability to move the surrounding application is more important.

The proof points should be operational. Enterprises should measure cost per successful workflow, completion rates, human intervention rates, unauthorized action attempts, approval frequency, recovery time, and the number of actions blocked incorrectly. For persistent agents, they should also track whether permissions and retained context remain appropriate as users, applications, and responsibilities change.

DevDay shows how quickly the market is moving beyond model selection. The next competition is over the agent environment itself: who supplies the runtime, maintains the context, governs the tools, and owns the interface through which people work. OpenAI now has a credible position across all four. Enterprises should take the productivity opportunity seriously while making sure convenience does not become architectural dependence.