OpenAI DevDay 2026: Choose the model from the vendor, keep your data in-house

At DevDay 2026, OpenAI introduced the Dots voice assistant, the GPT 6.1 Sol model with Ultrafast mode, Decisions API, Agents API, ChatGPT Spaces workspaces, and Marketplace in a single day.

The company also disclosed its audience: 1.2 billion people use ChatGPT each week (Latent Space analysis).

After these announcements, OpenAI sells the entire stack, from the model to the agent marketplace.

The rule for executives is clear: you can get the model from the vendor, but keep data, access rights, and the activity log in your own integration layer.

WAU (weekly active users) is the number of people who opened the service at least once during the week.

With a reach of 1.2 billion people, some of your employees are almost certainly already using ChatGPT from personal accounts and phones.

A procurement manager pastes a supplier contract into the chat, an analyst pastes a sales export, and an HR specialist pastes a salary table.

The security team does not know about it, and there is no request log.

With Spaces, storing work documents and managing collaborative projects in ChatGPT is even more convenient.

If the corporate tool is slower, an employee will move the work to ChatGPT, and a ban will not stop them.

People choose the tool that delivers results faster.

We call this metric TTU, time to use, and for ChatGPT it is measured in seconds.

With Agents API, a developer can assemble an agent in an evening that answers questions and calls a couple of functions.

Within a company, people expect more from an agent: stock levels from 1C, product cards from PIM, order statuses from ERP, and enforcement of the specific user's permissions.

You also need an audit log: who asked what, which data the agent saw, and what action it performed.

Without an audit log, security will shut down the project after the first data leak, and the person who launched it will be held responsible.

This is where most of the implementation budget goes, and OpenAI will not do it for you: the vendor does not know how your reference data is structured or who in the company may see margins.

The simpler the agent interface is for users, the more engineering work lies beneath it. The second risk is vendor lock-in.

If business logic lives in Decisions API, changing the model will require rewriting the processes.

For CIS companies, switching models is a practical necessity: access to OpenAI is limited, and data requirements mean some tasks must be assigned to GigaChat, YandexGPT, or Qwen and Llama within an in-house environment.

### Gateway between people and models LLM & Security Gateway accepts requests from employees and applications, masks personal data and trade secrets, maintains an audit log, and routes each request to the right model: GPT 6.1 Sol, Claude, GigaChat, or a local Llama.

Employees get speed, security gets logs, and the CFO gets a single bill for tokens.

To replace the model, the engineer changes the route in the gateway configuration. ###

Data access through MCP and RAG MCP (Model Context Protocol) - an open protocol through which an agent invokes tools.

Through MCP servers, we give the agent access to 1C, Pimcore, or Akeneo operations while respecting the user's permissions.

The protocol is supported by different model providers, so tools continue working when you change models. RAG is a search of the company's knowledge base before answering: the model responds based on your policies and price lists and cites the document. ### Agent actions through the bus

When an agent changes data, the change passes through the integration bus - Apache Kafka or Datareon - as an event from any other system.

The agent's action is checked by the same validation rules as a human action.

In PIM projects, we assign catalog routine to the LLM: generating filter facets, product card copy, and taxonomy classification.

The flow is the same for all these tasks: data comes from PIM, the model suggests an option, rules and a content manager validate it, and the integration returns the result to PIM.

When a stronger or cheaper model becomes available, we switch the route in the gateway and compare quality on the same product cards without rewriting the pipeline. Ultrafast in GPT 6.1 Sol may be useful here: batch processing thousands of cards is limited by response speed.

Testing on your data will show how much the mode speeds up work and what it costs; the figures from the presentation are not enough for such a decision.

At DevDay, OpenAI claimed the interface, office tools, and agent marketplace.

These services are cost-effective to use, but the vendor sets the terms of the deal.

The rule for executives: get the model and speed from the vendor, but keep data, permissions, logging, and routing in your own layer.

The test is simple: can you move a key AI process to another model within a week?

If not, you have bought a dependency, and OpenAI will set its price.