Evidence at a glance
Many Announcements, One Shift in the Control Plane
OpenAI introduced more than 20 updates at DevDay 2026 across ChatGPT, Codex, models, APIs, security, and new ways of working. The company presented ChatGPT as a shared surface for people and agents, where developers can launch native experiences and reach roughly 1.2 billion weekly users. The primary object here is not a new model in isolation, but a product structure that places users, agents, and developers at the same entry point.
That shifts the center of the event from whether a model is smarter to who controls the entry point and the complete work loop. GPT-6.1 Sol adds stronger capabilities for agentic coding, computer use, and professional work, while Codex can run locally, remotely from a phone, or in the cloud. Code review and security scanning then bring the output back into team processes. No single feature creates a platform, but when discovery, execution, checking, and distribution are placed on one product line, OpenAI is no longer merely supplying models to other systems.
Dots Turn a Request into an Ongoing Delegation
Dots is the clearest expression of this product direction. OpenAI describes it as an always-on agent able to take on ongoing responsibilities, learn what matters to a user, work on the user’s behalf, and remove important tasks from the user’s workload. The unit of interaction is no longer a question followed by an answer. It becomes a delegated relationship that can remain active, wait for events, and deliver outcomes over time.
Its deployment conditions reveal the other side of that capability. Dots is available to Pro and Business Premium users, while Enterprise, Edu, and Healthcare users can test it only after a workspace administrator enables it, with the beta off by default. Some access is also limited by market and plan. Persistent agency is therefore not being treated as ordinary chat functionality. For an enterprise, the question is not merely what an agent can do, but who can activate it, what it can access, when it must stop, and who is accountable when it is wrong.
Lower Prices and Higher Speed Change Automation Economics
The model updates give this work system more room to operate. OpenAI says GPT-6.1 Sol has input and output prices equal to one fifth of its standard pricing while delivering near-Astra intelligence. Ultrafast turns speed into an explicit product tier: GPT-6 Astra can reach up to 300 tokens per second in the API, Codex can be up to eight times faster, and the API can be up to six times faster. GPT-6 Astra Ultrafast is available in the API, with access through Pro 500 and Enterprise plans.
The importance of these figures is not simply that one request finishes sooner. They lower the threshold for assigning more steps to an agent. Lower token prices allow more complex workflows within the same budget, while lower latency makes code fixes, information handling, and repeated tool calls look more suitable for automation. Yet the cost model cannot stop at tokens. Retries, tool calls, human review, permission maintenance, recovery from mistakes, and business losses caused by an agent’s actions all enter the total cost as usage increases.
Codex Is Becoming a Schedulable Execution Environment
Codex is best understood here as a systems architecture change rather than an editor feature update. Developers can run tasks on a computer, operate them remotely from a phone, or send them to the cloud. Reusable development environments let tasks start quickly and give teams shared settings, approved configurations, and permissions. The CLI adds voice-based task initiation and steering, the /agents view helps users delegate and track multiple tasks, and the desktop app can summarize changes across projects, explore diffs, and answer questions about potential issues before feedback is sent to GitHub or GitLab.
This design changes the basic unit of development from an assistant inside an editor to a work environment that can be scheduled, handed off, and inspected. Automatic code review can take a first pass in the cloud while the developer is away. Codex Security Cloud can scan complete GitHub repositories on demand or on a schedule, continue checking new commits, investigate findings, remove duplicates, and prepare fixes in the cloud. The efficiency gain is real, but reusable environments also reuse configuration and permissions. Isolation, approval, and auditability therefore have to be designed into the execution layer rather than added as post-deployment patches.
Security Boundaries Will Determine How Far the Work Layer Can Go
OpenAI is also addressing the boundary that matters most to enterprise adoption. Private Intelligence includes Zero Data Retention with Private Safety Processing, allowing automated safety reviews without giving OpenAI personnel access to the underlying content. Private Inference is planned for preview this fall and combines confidential computing with strict, verifiable controls. For enterprises, these are not merely additional security labels. They attempt to answer where data appears while an agent operates, who can see it, and whether the controls can actually be verified.
An open ecosystem, however, is not the same as an open control plane. Developers gain a distribution surface inside ChatGPT and may build agents through the Agents API, plugins, and AWS-native resources. OpenAI’s collaboration with Amazon also brings core Agents API capabilities into Bedrock Managed Agents, allowing agents to run within AWS. The platform still controls review, permissions, and the user relationship, while enterprises still need to test whether Private Inference controls can be independently inspected. Code review, repository scanning, and reversible knowledge workflows are sensible early candidates. Payments, production releases, customer commitments, and irreversible data changes should remain under human control until approval, pause mechanisms, auditing, and accountability are explicit.