OpenAI DevDay 2026: From AI Assistant to Operating Layer
OpenAI DevDay 2026 was less about one breakthrough model and more about changing the role AI plays in everyday work.
The company announced more than 20 updates across ChatGPT, Codex, models, agents and developer infrastructure. The common thread is significant: OpenAI is moving from AI that responds to requests toward AI that can own work, operate software, collaborate with teams and remain active over time.
That transition matters more than any individual feature.
For product leaders, the question is therefore no longer simply, “Where can we add AI?”
It is becoming:
Which parts of our product, workflow and organization should humans own — and which can increasingly be delegated to persistent software agents?
Here are the most important launches from DevDay 2026, and the implications I see for products, individuals and companies.
1. Dots: OpenAI’s Move From Assistant to Persistent Agent
The most consequential announcement may be Dots.
OpenAI describes Dots as always-on agents that can work continuously on a user’s behalf. They run on their own cloud computer, can use a browser and connected applications, learn from feedback and can work across ChatGPT, Slack and Microsoft Teams.
Unlike a normal ChatGPT conversation, a Dot does not necessarily wait for the next prompt. It can continue working toward a goal, monitor information and bring completed work back for review.
OpenAI gives examples ranging from fixing software bugs and revising launch materials to updating research analyses and following enterprise sales opportunities.
Product leadership perspective
This changes the unit of product design.
Most software today is designed around sessions: a user opens an application, performs an action and leaves.
Agents introduce another model: delegation.
Instead of asking users to navigate workflows, product teams can increasingly ask:
What outcome should the user be able to delegate?
That sounds subtle, but it changes product architecture, UX, metrics and business models.
Success may become less about daily active users or time spent in the application and more about work successfully completed on the user’s behalf.
The opportunity
Products can evolve from tools into systems that actively help achieve outcomes.
A CRM, for example, does not necessarily need to wait for a salesperson to update opportunities. An agent could monitor conversations, identify missing information, prepare follow-ups and flag deals requiring intervention.
The strongest products may therefore combine proprietary data, domain-specific workflows and agent capabilities rather than simply adding another chat interface.
The risk
Delegation introduces an entirely new trust problem.
The more autonomy an agent receives, the greater the potential impact of incorrect actions, misunderstood intent or excessive permissions.
Product leaders will need explicit answers to questions such as:
- What can the agent do autonomously?
- What requires approval?
- What information can it access?
- How can users audit what happened?
- How can an action be reversed?
Agent UX will increasingly be permission UX.
OpenAI itself emphasizes configurable permissions, action review, activity tracking and safeguards for Dots.
Implication for individuals
The valuable skill may shift from producing every piece of work yourself toward defining outcomes, setting constraints and reviewing delegated work.
People who learn how to manage agents effectively could dramatically increase their personal operating capacity.
Implication for companies
Organizations need to start thinking about agents as a new type of digital workforce.
That means ownership, access control, auditing, governance and performance measurement cannot remain afterthoughts.
2. GPT-6.1 Sol: Frontier-Level Capability Becomes Cheaper
OpenAI also introduced GPT-6.1 Sol, positioned as a major upgrade to GPT-6 Sol.
The model is focused particularly on agentic coding, computer use and professional work. OpenAI says it delivers intelligence approaching GPT-6 Astra while costing one fifth of Astra’s standard input and output token prices.
Product leadership perspective
The important story here is not simply “the model got smarter.”
It is the capability-to-cost curve.
Capabilities that previously made sense only for high-value or exceptional workflows can become economically viable for ordinary product interactions when inference costs fall.
That expands the design space.
The opportunity
Teams can revisit AI use cases they previously rejected because they were too expensive, too slow or required too many model calls.
Agent workflows are particularly sensitive to this because a single user request can trigger many internal reasoning and tool-use steps.
Lower model cost therefore does not just make existing products cheaper. It can make entirely new product economics possible.
The risk
Falling model prices can quickly commoditize features built primarily around access to intelligence.
If your differentiation is “we use a powerful model,” that advantage is likely temporary.
Durable differentiation will increasingly come from workflow design, proprietary context, distribution, trust, domain expertise and accumulated user data.
Implication for individuals
Access to high-level reasoning and coding capability continues to become cheaper and more widely available.
The gap between people with access to advanced AI and those without it may shrink. The gap between people who know how to structure work around it and those who do not may grow.
Implication for companies
Revisit your AI business cases frequently.
A workflow that was economically unattractive six months ago may now be viable.
At the same time, assume competitors receive roughly the same model improvements you do.
The strategic question becomes what you build around the model.
3. Ultrafast: Latency Becomes a Product Feature
OpenAI introduced Ultrafast, a premium inference tier focused on speed.
OpenAI says Ultrafast can provide up to 8× faster token generation in Codex, reaching around 300 tokens per second, and up to 6× faster generation through the API. GPT-6 Astra Ultrafast is initially available in the API and selected ChatGPT Work and Codex plans.
Product leadership perspective
Latency is often underestimated in AI product strategy.
The difference between a response arriving in ten seconds and one arriving in one second does not merely improve the same experience.
It enables different experiences.
Near-real-time intelligence makes AI more viable inside interactive applications, collaborative environments, coding workflows and operational decision loops.
The opportunity
Product teams can begin designing AI that feels less like submitting a job and more like interacting with a live system.
That opens possibilities for real-time assistance, dynamic interfaces and faster multi-step agent workflows.
The risk
Speed has a cost.
Teams can easily pay premium inference prices for interactions where users would happily tolerate slower processing.
AI products increasingly need intelligent routing between fast, cheap and highly capable models rather than using one model tier for everything.
Implication for individuals
AI interactions become increasingly fluid and less disruptive to the normal flow of work.
Implication for companies
Latency should become an explicit product variable alongside quality and cost.
The best architecture may dynamically choose between them depending on the task.
4. Agents API and Computer Use: Agent Infrastructure Becomes a Platform
OpenAI expanded its Agents API with computer use.
Developers can now build agents that interact directly with software interfaces and combine those capabilities with multi-agent orchestration, tool search, tool calling and context management.
OpenAI operates the underlying infrastructure.
There is also an important enterprise distribution angle: OpenAI and Amazon announced Bedrock Managed Agents powered by OpenAI, allowing organizations to deploy related agent capabilities inside AWS environments.
Product leadership perspective
This is a meaningful abstraction shift.
Developers previously needed to assemble much of the agent stack themselves: model orchestration, tools, context handling, execution environments and reliability logic.
More of that infrastructure is moving into managed platforms.
The result is predictable: building an agent becomes easier.
But building a valuable agent business does not automatically become easier.
The opportunity
Smaller teams can build sophisticated automation products that previously required substantial infrastructure investment.
Legacy software also becomes more accessible to AI. If an agent can operate a graphical interface, companies do not always need a perfect API before automation becomes possible.
The risk
Computer use introduces fragility.
Interfaces change. Buttons move. Authentication fails. Instructions conflict. External websites may contain hostile or misleading content.
For critical workflows, “the agent can click the interface” should not be confused with “the workflow is reliable.”
Implication for individuals
More repetitive workflows across separate applications can be automated without waiting for every application vendor to build native integrations.
Implication for companies
Expect a growing architectural choice between API-based automation and interface-based agent automation.
Use computer interaction strategically, but keep reliability, observability and fallback paths in the design.
5. Decisions API: Turning AI Into a Decision Component
OpenAI introduced the Decisions API, which applies Luna’s intelligence to constrained questions with predefined possible answers.
Developers can provide text or images and receive structured decisions that can be used for classification, routing or determining an agent’s next action.
Product leadership perspective
This may look less exciting than an autonomous agent, but it could be extremely useful.
Many enterprise problems are not open-ended generation problems.
They are decisions:
Is this request urgent?
Which workflow should handle this case?
Which category does this document belong to?
Should the agent continue or escalate?
Making those decisions reliable, fast and inexpensive is essential if agent systems are going to operate at scale.
The opportunity
AI can increasingly sit inside operational systems without exposing a conversational interface at all.
Some of the most valuable AI products may therefore be nearly invisible to the end user.
The risk
Automated decisions can silently scale mistakes.
Teams need monitoring, confidence thresholds, escalation logic and periodic evaluation rather than assuming model output is automatically correct.
Implication for individuals
Users may experience better routing and automation without necessarily interacting directly with an AI assistant.
Implication for companies
Treat AI decision components like production infrastructure: measure them continuously and define where humans remain accountable.
6. Codex Moves From Coding Assistant to Software-Engineering System
OpenAI announced several significant Codex upgrades.
Codex in the cloud allows developers to run coding tasks remotely from different devices using reusable development environments.
The refreshed Codex CLI adds voice control and a multi-agent view for delegating and tracking parallel tasks.
A new Code Review experience helps developers inspect changes, summarize diffs and identify potential problems before pull requests or merge requests are approved.
Codex Security Cloud can scan repositories continuously, investigate findings, remove duplicates and prepare fixes in the cloud.
Product leadership perspective
Coding agents are moving up the abstraction ladder.
The first generation helped developers write individual functions.
The next generation can own increasingly complete engineering tasks: investigating an issue, modifying code, testing the implementation, reviewing changes and monitoring security.
That changes software economics.
The opportunity
Product organizations may be able to dramatically increase the number of experiments, fixes and incremental improvements they can ship.
The constraint may move from engineering throughput toward deciding what is worth building.
That increases the value of strong product judgment.
The risk
More code is not automatically more product value.
AI can make it extremely inexpensive to produce technically functional features that nobody needs.
Technical debt can also scale faster when generation becomes cheap.
Implication for individuals
Engineers increasingly become orchestrators and reviewers of parallel implementation work rather than authors of every line of code.
Product managers with technical fluency may also be able to prototype much more independently.
Implication for companies
Engineering productivity metrics will need to evolve.
Lines of code, commits and even feature throughput become weaker proxies for performance when software production becomes abundant.
Customer impact becomes more important, not less.
7. Plugins Become a Distribution Platform Inside ChatGPT
OpenAI significantly expanded its plugin ecosystem.
Developers can now create plugin extensions with dedicated sidebar experiences, interactive panels and custom file viewers directly inside ChatGPT.
OpenAI is also improving plugin creation, submission and discovery, including Plugin Creator and recommendation systems.
Supported plugins can additionally run inside ChatGPT Sites, while support for MCP Events enables connected applications to trigger automations when external events occur.
Product leadership perspective
This points toward ChatGPT becoming more than an AI application.
It is increasingly becoming a distribution surface for other software products.
That creates echoes of previous platform transitions: the browser, mobile app stores, Slack apps and cloud marketplaces.
The opportunity
A startup may no longer need to convince users to adopt an entirely new destination.
It can potentially bring its product into an environment where users already work.
OpenAI says ChatGPT now reaches 1.2 billion weekly users, making distribution inside ChatGPT strategically interesting for developers.
The risk
Platform distribution creates platform dependency.
If discovery, ranking, permissions and interaction patterns are controlled by ChatGPT, developers may gain distribution while losing part of the direct customer relationship.
The familiar platform question returns:
Are you building a company or a feature inside someone else’s ecosystem?
Implication for individuals
Users may increasingly access specialist software without leaving their primary AI environment.
Implication for companies
Plugin strategy may become as important for some businesses as mobile or API strategy.
But companies should preserve channels and customer relationships they control themselves.
8. ChatGPT Space and Pages: AI Enters the Collaborative Workspace
OpenAI introduced ChatGPT Space, a shared environment where teams, ChatGPT and agents can work from common knowledge.
Within Spaces, Pages provide collaborative documents where people and AI can write, research, create visualizations, generate images and work together.
OpenAI is also introducing collaborative slide creation, allowing multiple people and agents to work on a presentation and export it to PowerPoint or Google Slides.
Product leadership perspective
This is OpenAI moving directly into the collaborative-work category.
The traditional enterprise stack separates communication, documents, presentations, project management and AI.
OpenAI is trying to make AI the connective layer across all of them.
The long-term battle may therefore not be about “the best chatbot.”
It may be about where organizational context accumulates.
The opportunity
If the AI shares the same workspace, documents and project context as the team, users spend less time repeatedly explaining context.
That makes agent performance more useful because the system has persistent organizational knowledge.
The risk
The more knowledge that accumulates in one platform, the greater the consequences of poor information architecture, permissions or vendor lock-in.
Organizations also need to decide which information should become persistent AI context and which should not.
Implication for individuals
Work may become less fragmented across conversations, documents and AI tools.
Implication for companies
Knowledge architecture becomes part of AI strategy.
Companies with clean permissions, structured information and clear ownership will have an advantage over organizations whose knowledge remains scattered across hundreds of disconnected systems.
9. Shared Teams, Tasks and ChatGPT Inside Slack and Teams
Business and Enterprise customers can create teams that share Pages, slides, spreadsheets and plugins.
Teams can also delegate recurring work to shared tasks triggered by schedules or events such as new emails or messages.
OpenAI additionally announced @ChatGPT integrations for Slack and Microsoft Teams, allowing teams to involve ChatGPT directly in existing conversations and use connected company tools according to permissions.
Product leadership perspective
AI adoption often fails because the tool sits outside the workflow.
Embedding AI inside Slack and Teams attacks that problem directly.
Instead of asking everyone to develop a new work habit, the AI enters an existing one.
The opportunity
This dramatically lowers the activation barrier for enterprise AI.
The interaction model becomes as simple as bringing another colleague into the conversation.
The risk
Ease of access can create uncontrolled proliferation.
If employees can invoke agents everywhere, companies need clear rules around data access, actions, accountability and escalation.
Implication for individuals
Using AI increasingly becomes part of normal collaboration rather than a separate activity.
Implication for companies
AI governance needs to move from tool approval to workflow governance.
The important question is no longer only “Can employees use ChatGPT?”
It is “What can ChatGPT do inside our organization?”
10. Meetings Plugin: Meetings Become Structured Input for Agents
OpenAI’s new Meetings plugin can capture meetings, create personalized notes and action items, and save them into ChatGPT Space.
Those outputs can then feed directly into follow-up work such as updating a project plan or drafting communications. OpenAI says the meeting audio is deleted after the notes are generated.
Product leadership perspective
Meeting transcription itself is not new.
The interesting part is what happens after the transcription.
If meeting decisions automatically update plans, trigger work and inform agents, meetings stop being isolated records and become events in an operational system.
The opportunity
A large amount of organizational information currently dies in meetings.
Connecting conversations directly to execution could remove considerable coordination overhead.
The risk
Recorded meetings create understandable privacy and behavioural concerns.
Employees need clarity about what is captured, retained, shared and acted on.
Implication for individuals
Less manual note-taking and follow-up.
But individuals also need to become more conscious that spoken decisions may become persistent operational context.
Implication for companies
Meeting AI should be treated as part of the company’s knowledge and governance architecture, not simply as a productivity add-on.
11. Private Intelligence: Enterprise AI Moves Closer to Sensitive Workloads
OpenAI introduced Private Intelligence, including Zero Data Retention with Private Safety Processing and a preview of Private Inference.
The goal is to enable organizations to use advanced models while providing stronger controls around how sensitive information is processed.
Product leadership perspective
AI capability has rarely been the only blocker for enterprise adoption.
Trust, security and data handling are often equally important.
As models become capable enough for high-value professional workflows, privacy infrastructure becomes a core product capability rather than a compliance appendix.
The opportunity
Stronger privacy guarantees can open workloads in finance, healthcare, legal services, engineering and other environments where sensitive information previously limited adoption.
The risk
“Private AI” can easily become a marketing phrase.
Enterprise buyers should evaluate the exact technical architecture, retention policy, access controls and contractual guarantees rather than relying on labels.
Implication for individuals
More sensitive professional work may become eligible for AI assistance.
Implication for companies
Security architecture will increasingly determine how much value organizations can safely extract from frontier models.
12. Sign in with ChatGPT: Your AI Subscription Starts Following You
OpenAI also announced Sign in with ChatGPT.
Users can authenticate with participating third-party products and, in some cases, use portions of their ChatGPT plan allowance across partner tools.
OpenAI named partners including Devin, Notion and Vercel among the initial integrations.
Product leadership perspective
This could become strategically important.
Identity, distribution and model usage are beginning to converge.
If users bring their existing AI subscription into third-party products, developers may be able to offer sophisticated AI functionality without owning the entire billing relationship for inference.
The opportunity
Lower friction for both adoption and monetization.
The risk
It gives OpenAI an increasingly central position between applications and their users.
Companies need to consider what happens to their economics and customer relationships if AI access becomes controlled by an external subscription platform.
Implication for individuals
One AI subscription may increasingly unlock functionality across multiple products.
Implication for companies
AI entitlement could become another platform dependency to model alongside cloud, app stores and payment infrastructure.
13. OpenAI Marketplace: AI Procurement Becomes an Ecosystem
OpenAI also introduced an OpenAI Marketplace for eligible enterprise customers.
Companies can potentially apply part of their existing OpenAI commitment toward approved partner software. Initial partners span design, customer experience, legal, cybersecurity and infrastructure.
Product leadership perspective
This is an ecosystem move as much as a procurement feature.
OpenAI is creating a mechanism through which enterprise AI budgets can flow toward third-party applications.
That could make ChatGPT and OpenAI infrastructure not just another vendor in the technology stack, but a channel through which other vendors are purchased.
The opportunity
For B2B AI companies, marketplace inclusion could shorten procurement cycles and unlock enterprise distribution.
The risk
Distribution through a dominant platform can also increase dependency on its commercial terms and approval processes.
Implication for individuals
Little changes immediately.
Implication for companies
For AI vendors, platform partnerships may become a meaningful go-to-market channel.
For buyers, AI procurement may gradually consolidate around fewer strategic platform relationships.
The Bigger Product Shift: From Software Tools to Delegated Outcomes
Looking across the DevDay announcements, three transitions stand out.
First, AI is becoming persistent.
Dots, recurring tasks and event-triggered automations mean AI increasingly continues working after the conversation ends.
Second, AI is becoming operational.
Computer use, the Agents API, Codex and plugins allow models to interact with systems rather than merely explain what a human should do.
Third, AI is becoming collaborative infrastructure.
Spaces, Pages, shared tasks, Slack, Teams and plugins move AI into the places where organizations already coordinate work.
Put together, these changes suggest that we are moving beyond the era of the AI assistant.
The emerging product is closer to an AI operating layer sitting between people, software, organizational knowledge and work.
What Product Leaders Should Do Now
The temptation after an event like DevDay is to create a list of features to copy.
I think that misses the larger opportunity.
Instead, product leaders should revisit four fundamental questions.
What work can users delegate rather than perform?
Look for complete outcomes, not individual AI features.
Where does your differentiation live if intelligence becomes cheap?
Model capability will continue to spread. Proprietary workflow, context, distribution and trust become more important.
What should an agent be allowed to do?
Design permissions, approvals, observability and recovery alongside the core experience — not after launch.
What happens to your product when users interact through an agent instead of your interface?
This may be the most uncomfortable question.
If AI increasingly mediates how people interact with software, some traditional UI, navigation and engagement advantages become less valuable.
The companies that benefit most may be those whose products remain valuable even when nobody clicks through their interface.
Final Takeaway
OpenAI DevDay 2026 was not primarily an announcement about a better chatbot.
It was an announcement about a different computing model.
AI is moving from something we consult toward something we delegate to.
That creates an enormous opportunity to remove coordination work, automate workflows and give individuals capabilities that previously required entire teams.
It also shifts the product leadership challenge.
When software can reason, act and continue working independently, building the capability is only half the job.
The harder questions become:
What should it do?
What should it know?
What should it be allowed to change?
And where should humans remain firmly in control?
Those questions will increasingly define good product leadership in the agent era.
Discover more from The Tech Society
Subscribe to get the latest posts sent to your email.