OpenAI Product Head Thibault Sottiaux on Why the World Seems Ready

OpenAI’s software engineering tool, Codex, has long been associated with Thibault Sottiaux, the engineer known for resetting token limits whenever the product hits a major growth milestone. Now, OpenAI aims to deliver that same sense of achievement to a wider audience through ChatGPT Work, a platform designed to help white-collar professionals harness AI agents. We recently spoke with Sottiaux about the product’s challenges, but we wanted to share a broader selection of that conversation with our readers. Below is a lightly edited and condensed version of our interview, covering strategies for winning over skeptics, treating discovery as a core design philosophy, and the true cost of intelligence.

TechCrunch: I understand you’re part of the technical staff. Can we accurately describe you as the product lead for both Codex and Work?

Sottiaux: I oversee all core products, which includes the API, agent infrastructure, enterprise solutions, and the entire ChatGPT suite—encompassing both ChatGPT Work and the classic version—as well as everything related to Codex.

Q: Is it correct to assume that you report directly to Greg Brockman?

Sottiaux: Yes, Greg. I often joke that everyone ultimately reports to Greg.

Q: What makes ChatGPT Work so critical to OpenAI’s strategy?

Sottiaux: Our goal was to democratize the power of coding agents. This initiative takes technology originally designed for technical experts and repackages it to be safe, delightful, and accessible on mobile and web platforms. By launching it as part of the Plus plan at just $20 a month, we aim to provide incredible value to the broadest possible audience. That accessibility and utility are exactly what we set out to achieve.

Q: Analysts suggest there is a significant economic driver here: OpenAI needs to secure the application-level relationship with users.

Sottiaux: The more value and utility we generate for users, the more willing they are to pay for that service. This has always been our perspective on ChatGPT. When users realize the immense utility they receive, paying $20 a month feels like a no-brainer because the value far outweighs the cost.

Q: How significant is this technology in shifting public opinion toward AI?

Sottiaux: We believe we have a responsibility to bring everyone along with this technology. The question is how to achieve widespread adoption. Codex was initially built for a forgiving technical audience, allowing us to release capabilities early. Now, with the technology reaching a mature level, we feel it is the right time to diffuse it to a much broader audience. This goes beyond simple writing assistance or personal advice; ChatGPT can now autonomously handle complex tasks in a way that is both safe and delightful. Our mission at OpenAI is to ensure everyone benefits from this progress.

TC: How do you approach the transition between letting the model take the lead and providing users with buttons or a traditional interface?

Sottiaux: Our core objective is to build extremely capable models and then find the simplest, most delightful way to integrate them into users’ lives. To do this effectively, we must step back and let the model express itself, allowing users to harness its full utility. This means maintaining a minimal product surface with delightful simplicity and engaging with humans in a natural way. We recently launched ChatGPT Voice, which has seen significant growth, because talking to it is super natural. We are having a conversation right now. As technology progresses, interactions will become increasingly natural, adapting to humans rather than forcing users to learn how to use the application.

TC: Is the model ready? I read that Ethan Mollick, a Wharton professor studying these tools, noted that ChatGPT Work aims for “magic,” while Claude Cowork relies on A/B tests and user choices. Do you believe workers are generally ready for that level of autonomy?

Sottiaux: The world appears ready, which is why we’ve seen incredible adoption. We recently announced hitting 20 million users. We launched it in a way that is simple yet powerful, uncompromising in quality, and easy to use.

TC: My mind struggles to conceive of building a product meant to do everything. Do you focus on a minimum viable product or discrete problems? What metrics do you evaluate?

Sottiaux: Product development is also an exercise in discovery. As we push the boundaries of model capabilities, we discover what the models can do and lean into their strongest areas to build great products. This element of discovering capabilities is also magical and fun for us. For instance, GPT 5.6 represented a step up in general work capabilities, such as processing large documents, generating high-quality slides and reports, and conducting deep research—tasks typical for professionals. We then focus on these areas, capture feedback, and continue to improve. This iterative deployment involves learning from the community and real-world usage to continuously enhance the product.

TC: There is a noticeable gap between the cost of the Plus plan and the amount of tokens consumed. How do you view this dynamic? Should users or CFOs be concerned?

Sottiaux: We work daily to push the frontier of efficiency. With the announcement of major price cuts for Luna, offering an 80% reduction, we have implemented a permanent price correction. Frontier capabilities are becoming increasingly affordable over time, and this trend will continue. Our goal is to include more utility within the same price point. While users can always pay more for additional capabilities, today’s features should allow users to achieve the same results with less spending six months from now.

TC: What do you say to users who are stressed about granting the model access to their email or iMessages, features you just rolled out?

Sottiaux: It is crucial to select models that are safe and aligned. A significant portion of our investment goes into our safety stack, our safety approach, and publishing honest benchmarks on these metrics. Our models are world-class in these areas, ensuring that user data and interactions remain secure.