Arnaud Fournier is a Europe engineering manager at Openai, attached to the French office of the company. It details the strategy of the start-up which now claims 3 million professional users.
JDN. OPENAI today claims 3 million professional users. In a context where competition is intensifying, in particular with Google and Anthropic, what is your strategy to maintain, even strengthen, your advance on the market for professional AI assistants?
Arnaud Fournier. We actually have 3 million professional users today, compared to 2 million in February, which testifies to remarkable growth. We arrived a little later on this segment, Chatgpt having been launched two and a half years ago. The demand has been massive and naturally evolved from consumer uses to professional applications. Companies express an increasing need for integration into their information systems.
You have recently announced Deep Research integrations with Github, HubSpot, Linear in particular. What do you think will be the priority use cases to send for professionals?
Integrations with Github, Hubspot and Linear respond to the growing demand of our customers who wish to deepen their chatgpt uses. Take the example of developers: with Deep Research, they can now obtain a precise analysis when receiving a brief. The tool makes it possible to quickly identify the code areas requiring intervention, to orchestrate the first phase of work and to make the developer more effective in the architecture of its solutions. The GitHub and Linear connectors allow you to interact directly with the knowledge bases and to carry out more analytical work.
The HubSpot connector, for example, will mainly target sales teams using CRMs. Our long -term vision is to make Chatgpt a universal platform, capable of serving each profession and each individual in their professional interactions.
The MCP communication protocol has aroused growing interest for a few months. Is it a fundamental trend called to intensify permanently?
What we see is that we have gone from with the generative AI of a revolution to a continuum of revolution with agentic AI today. We made a lot of announcements a few months ago with the tools for the developers, our SDK Agent, the API Response, which allow everyone to create their own agents. And to be able to go even further, these agents must be able to interact with tools and systems. And we, of course, create these connectors directly for lots of tools, and we will continue to do it, but we also want to give the developers to create their own connectors in their sources systems, sometimes to which we do not have access, because it can be systems at home that are separated from the Internet. And so, by integrating MCP, we also unlock this ability for everyone to integrate with these tools, and we see that this protocol is becoming prevailing in the industry.
Today, we observe the emergence of specialized agents capable of performing tasks in the background, independently, as is the case in projects such as Codex or Deep Research. Are we going to systems built around multiple invisible but active agents, each dedicated to a specific function?
These agents actually work in the background, but whenever there are tasks presenting a certain risk – including in Codex or in the operator – validation is systematically requested from the user. The important thing is to identify the use cases for which we have a sufficient level of confidence. For example, getting information on the Internet with Deep Research does not imply a very high level of risk. On the other hand, as soon as you have tasks that actually run in systems, there is absolutely necessary to maintain this human intervention.
“We want to give the developers to create their own connectors in their sources systems”
All these tools allow this interaction with the user by asking: “I would like to launch this command, do you allow me?” It is this human -machine interaction that we want to integrate into Chatgpt to ensure the alignment between this agent – which can save us time – and the need to not lose more by tasks that could be wrong. This is the error that can be observed when people try to do this in total autonomy or try to go faster than music.
Codex is now available in two versions: one integrated with chatgpt and connected to Github, the other in command line via Codex CLA. Do these two tools respond to very distinct use cases, or are we talking about the same type of tasks, but applied to different technical contexts?
These two products are very complementary and reflect the philosophy of OpenAi: on the one hand, make these AI tools accessible to the general public-whether professional developers or people who code the weekend-and on the other hand more advanced tools to those who have personalization needs or who wish to integrate them into their products and services. Users with strong expertise, well connected to these tools and databases, can thus create products with high added value thanks to underlying OPENAI technologies-our reasoning models, our multimodal models and all the tools for developers that we have created and launched.
The phenomenon of Vibe Coding is gaining popularity. But faced with this progress, do you still think that there will remain a space for creation specific to human developers in 5 or 10 years? Or are we going to complete software development automation?
Vibe Coding is not made to encourage bad code. Our objective at OpenAi is to equip tool developers allowing them to delegate tasks with low added value. Take concrete examples: unit tests and documentation are often carried out at least. A developer will generally spend little time properly documenting a function or writing exhaustive tests, while these aspects are crucial for the maintenance of the code.
“We work to develop models with larger context, including models of reasoning”
Our mission is not to replace the developers, but to provide them with means of being more effective. What I observe daily is that companies that have integrated these new coding capacities recruit more and work more efficiently. For example, a tool that automatically generates technical documentation or that produces test cases can release time for higher added value tasks such as software architecture or innovation.
Several IA research projects are focusing on the infinite context, capable of ingesting the complete history of a project or a deposit without prior division. Is this a direction that you also explore at Openai?
On this subject of the long context, we have recently announced the release of GPT-4.1, which is a non-grade model but which precisely has a much more extensive context. We have a million tokens for GPT-4.1, which is already a significant form of advance. Today, we are working to develop models with larger context, including models of reasoning.
You mentioned the advances of GPT-4.1 on the long context. Beyond the increase in the size of the context, how do you work concretely on technical challenges such as “Lost in the Middle” which still affects long context models in business?
We actually carry out a lot of research on this subject. When GPT-4.1’s announcement, our work was precisely not to get out of long-context models that would not work properly. The work we have done on GPT-4.1 allowed us to post-training our models, that is to say to improve their training specifically on this problem. The performances of GPT-4.1 are really excellent on the question of “Lost in the Middle”.
“We are convinced that we still have a lot to bring both the size of the context, but also on the quality of rendering”
Of course, this is a subject on which we continue to work. The challenge, when you are at the forefront of innovation, is that it is always difficult to predict what will come. We know that we work, we know that this area will evolve. What direction will it take? How will it go? It’s always difficult to say. I do not necessarily have certainties to share, except that we are actively working on the subject and that we are convinced that we still have a lot to bring in this area – both on the size of the context, but also on the quality of rendering.
