OpenAI Is Expanding the Business Around the Model

Four DevDays reveal a broader offer: intelligence, execution, distribution, shared work and purchasing access.

The strategic news from OpenAI’s developer conferences is the widening scope of what the company sells. Its first DevDay expanded models and application-building tools. Its 2026 event reaches into shared work, persistent agents, application distribution and enterprise purchasing. A business evaluating OpenAI now needs more than a model comparison.

My reading: OpenAI is assembling several ways to participate in the same piece of work—from supplying intelligence to hosting execution, organizing collaboration, and influencing how software gets discovered and purchased. That is an interpretation of its announced products, not evidence of a predetermined master plan or an inevitable market outcome.

The opportunity is fewer components to assemble. The tradeoff is a broader dependency to manage. Evaluate the complete working arrangement: execution, information, distribution, cost and exit. The rest of this analysis separates the documented progression from what businesses should infer—and what they still need to test.

When I read a developer-conference announcement, I’m interested in the demo. I’m more interested in the purchasing decision behind it. Does the company want me to rent intelligence, build on its tools, reach its audience or move my team’s work into its product? Those choices create different obligations long after the keynote ends.

Put a hypothetical sales-proposal workflow on the table. It needs approved product information, customer context, a draft, review and a record of what was sent. One supplier might provide only the language model. Another might supply the work environment, integrations, agent, document and purchasing agreement. I would assess those offers differently even if their first drafts looked identical.

That is the tension I see in OpenAI’s progression. Integration can remove work that nobody on a small team wants to own. It can also make the team’s information and routines harder to separate from the service. Neither outcome follows automatically. I want to know exactly what the integration saves and what it makes expensive to change.

So I would read these DevDays as a sequence of expanding offers, with some revisions along the way. The useful question for a leadership team is where that expansion overlaps its own priorities. A model upgrade can justify a benchmark. A new execution environment calls for an architecture decision. A new route to customers calls for a distribution strategy. Treating all three as “more AI” obscures the decisions that matter.

A simple account would say OpenAI began as a model supplier and is becoming a software platform. The chronology is less tidy. ChatGPT Enterprise launched in August 2023, before the first DevDay, with administrative controls and business-data protections. The application business and enterprise relationship were already present.

The better interpretation is accumulation. Successive events add ways to build, distribute and operate software around the models. Some additions survive; others change direction. That distinction matters because a strategy can become broader while individual products become less durable.

At the November 2023 DevDay, OpenAI introduced lower-priced model access, a larger context window, new multimodal capabilities and the Assistants API. That API combined tools and persistent conversation threads to help developers construct assistive applications. These were building blocks for turning model capability into a product someone could use.

The business implication was practical. A developer could spend less effort assembling common functions and more on the application’s specific job. Yet the developer still needed to decide what a correct outcome meant, which information belonged in the workflow and how the customer would encounter the service. A more convenient API did not settle those questions.

The distribution ambition was visible at the same event. OpenAI introduced custom GPTs, allowed people to share them and announced a coming store. The accurate historical claim is that a storefront was announced, not that every promised commercial feature arrived that day. Already, OpenAI was courting both professional developers and people packaging expertise without writing a conventional application.

The 2024 DevDay program emphasized Realtime, vision fine-tuning, prompt caching and model distillation. The Realtime API launch offered a speech-to-speech interface in public beta. Prompt caching offered discounted processing for recently seen inputs. Distillation connected stored examples, evaluation and fine-tuning to help developers adapt smaller models to particular tasks.

Read together, these announcements addressed different reasons an application might be difficult to operate: interaction delay, repeated processing and the cost of using a large model for every request. This is the less theatrical work of a developer platform. It gives builders more ways to choose the balance between capability and operating expense.

For a business customer, the lesson is to assess the workload before celebrating a general price improvement. A discount on repeated inputs helps only where that pattern occurs. A smaller adapted model needs to meet the task’s quality standard. A natural voice interaction still needs a reliable path to the requested outcome. Lower component costs are an input to a business case, not the business case itself.

In October 2025, AgentKit brought together visual workflow design, connector management, embedded chat interfaces and evaluation capabilities. At the same DevDay, apps in ChatGPT and the Apps SDK offered developers a way to make their products available within conversations. App submissions opened in December, with review before publication.

Those are two distinct offers. One helps developers produce an application. The other helps customers encounter it. When the same company supplies both, builders should examine each on its own merits. A useful development toolkit does not establish that its distribution channel will deliver customers at an acceptable cost. A promising channel does not require placing every part of a product inside it.

The current AgentKit page also documents a reversal: a June 2026 update says its Agent Builder and Evals products will become unavailable on November 30, directing users toward code-based or workspace alternatives. That notice concerns those named products, not the practice of evaluation. It is concrete evidence for keeping workflow definitions and test cases recoverable outside a particular interface.

The 2026 recap brings ongoing agents, shared work and developer extensions into the same event. Its strategic significance is the connection between them. An application can become a place to work, a capability an agent calls and an experience distributed within a larger workspace.

Three levels help explain the change. For an individual, current ChatGPT Space supports pages where people and agents work with shared context. For a team, the same product describes real-time editing and comments. For organizations evaluating an AWS preview, Bedrock Managed Agents introduces a preview of OpenAI-powered agents operating inside AWS with AWS identities and permissions. These are different deployment decisions, not interchangeable ways to buy a smarter answer.

The strategic opportunity is continuity: less re-explaining a task between its conversation, files, execution and review. The corresponding obligation is continuity of control. Who owns the instructions? Which record is authoritative? Who can change the access? A workspace earns its place when those answers remain clear as more people and agents participate.

Current plugin extensions let developers provide sidebar applications, conversation panels and custom file viewers. That goes beyond placing a response inside a chat bubble. It offers application makers a more persistent presence within ChatGPT. The documentation also records rollout differences: some web extensions for Free and Go are still coming, and composer mentions are desktop-only.

For builders, the opportunity is to meet a customer near the moment a task arises. The tradeoff is dependence on another product’s discovery, permissions and interface conventions. The right experiment measures whether customers return to complete a meaningful job. Installation counts alone would not establish that the channel is working.

Infrastructure shows a different kind of expansion. The September 10 Agents API announcement describes a managed system based on Codex’s agent coordination software, with options for OpenAI-hosted, customer-operated or partner environments. The AWS preview says agent runtime and model inference remain inside AWS. These options complicate a simple story about everything moving into one vendor’s cloud.

The important distinction is between adopting a vendor’s intelligence and adopting all of its operating choices. Managed coordination could reduce engineering work; it also creates an interface and lifecycle dependency. Keeping an execution environment in a familiar cloud could simplify some governance questions without making the overall application portable. Buyers should ask which parts move together and which can actually be replaced.

OpenAI Marketplace’s current terms describe a qualified program: eligibility depends on the agreement, product and program terms. Customers contract with and pay the partner directly, while OpenAI reconciles the applicable commitment. This is not a general promise that any software purchase can consume any OpenAI budget.

The strategic inference is narrower and more useful. Existing commitments may influence which products a buyer seriously evaluates. For a software company, procurement compatibility could matter alongside technical integration. For a customer, an apparently convenient use of committed funds still needs comparison with alternatives, including the cost of the next renewal after the initial commitment is exhausted.

Another route extends the subscription outward. OpenAI’s plan-usage documentation allows eligible Plus and Pro users to use their ChatGPT allowance in participating apps, subject to limits. The app can charge separately. Signing in does not itself grant access to conversation history or memory. This separates identity, usage entitlement and data permission—three things a buyer should never assume are the same.

A post-DevDay development adds another commercial channel. On October 5, OpenAI announced a visual advertising format and expanded measurement. The initial visual-format test is planned for later this month in the US. OpenAI says ads remain labeled, separate from generated images and independent of answers. Those are stated product commitments, not conclusions from an independent audit.

The editorial implication is that application discovery, enterprise purchasing, subscriptions and advertising are becoming parallel ways to participate in OpenAI’s ecosystem. Their incentives differ. Developers should investigate how each channel works before treating “being on the platform” as a single growth strategy. Business customers should keep product suitability separate from the convenience of how a product is offered.

Phase 1: Map one complete job. Choose a workflow with a recognizable beginning, finished artifact and accountable owner. A proposal, customer-support resolution or internal planning brief is more useful than a general ambition to “adopt agents.”

Phase 2: Compare operating arrangements. Test a bounded version with managed services and identify what your team would otherwise need to supply. The purpose is to discover a useful division of responsibility, not to prove that buying or building is always superior.

Phase 3: Make the commercial decision follow the evidence. Bring engineering, the workflow owner and procurement into the same review. A successful demonstration should produce a specific proposal: what to adopt, what to retain and what conditions trigger reconsideration.

The key success factors are straightforward:

ROI considerations: The proposed measure is cost per accepted outcome, with acceptance defined before the trial. A faster intermediate step matters only if the completed process benefits. Record three quantities:

Competitive implications: For builders, the defensible position should come from the customer problem, specialized information, trusted execution or distribution relationships they can sustain. For buyers, the attractive arrangement is the one that improves a real process while preserving enough control to change it. Neither judgment can be read directly from a feature count.

For the next planning cycle, treat OpenAI as a supplier offering several potentially connected roles. Build a proposal around the roles your organization actually needs: intelligence, development tools, managed execution, collaboration, distribution or purchasing access. Give each a reason to be included. Shared branding is not a business requirement.

The boardroom assumption to challenge is that choosing a model settles the AI strategy. Model quality remains important, but the operating arrangement determines where information accumulates, how work is reviewed and what a change of supplier would require. Those are management decisions that cannot be delegated to a leaderboard.

Run a bounded evaluation with a real workflow owner. Establish what success means, preserve the starting cost and quality baseline, and compare the complete result. Record what your team stops doing, what new work it takes on and what evidence would justify expanding the arrangement. Keep announced features outside the acceptance case until the people responsible can use and verify them.

My conclusion is conditional: OpenAI’s expanding offer can be valuable where it removes real coordination and operating work. That value should be demonstrated at the same level at which the dependency is created. Bring this question to the next meeting: Which responsibilities are we deliberately buying—and what would it take to reclaim them?