How OpenAI and Databricks are working together

Summary

• GPT-5.5 launch: OpenAI's latest model delivers frontier reasoning and problem-solving capabilities, now integrated into Databricks Data and AI platform.
• Enterprise AI convergence: Databricks' partnership with OpenAI solves the fundamental challenge: how to apply frontier AI to enterprise data and operational use cases.
• Data + AI alignment: Enterprise deployments require securing data access, maintaining compliance, and enabling all employees—not just developers—to leverage frontier intelligence.

How OpenAI and Databricks are working together

OpenAI CRO Denise Holland Dresser in conversation with Databricks Co-founder Patrick Wendell discussing the launch of GPT-5.5, impact on joint customers and how builders and non-technical teams can harness the power of this latest model directly in Databricks’ platform.

Chapters

• 0:00 Introduction to AI partnership
• 1:39 Addressing enterprise challenges
• 3:35 Scaling AI to production
• 5:07 Security and control frameworks
• 6:25 Performance of GPT-5.5
• 8:44 Office QA benchmark results
• 10:53 Future of enterprise AI

FAQs

What makes GPT-5.5 significant for enterprise AI?

GPT-5.5 delivers improved reasoning and problem-solving capabilities compared to earlier models. Integration into Databricks enables enterprises to apply this frontier intelligence directly to proprietary data and operational decisions within governance frameworks.

What challenges do enterprises face deploying AI?

Enterprises must solve two critical challenges: providing frontier AI access while managing data security and enabling all employees—not just engineers—to leverage AI. Databricks addresses both by combining OpenAI's models with unified governance and data access.

How does the OpenAI-Databricks partnership help customers?

The partnership combines frontier intelligence (GPT-5.5) with enterprise data and AI infrastructure (Databricks). Customers can build agents and applications that reason over proprietary data securely, unlocking business value that neither platform provides alone.

How can non-technical teams use this partnership?

Builders and non-technical users can harness GPT-5.5 directly through Databricks' platform without writing code. This democratizes access to frontier AI capabilities across business functions, enabling every employee to benefit from intelligent automation.

Full transcript

0:00 Introduction to AI partnership

I'm excited to be joined today by, co-founder of Databricks, to talk about the launch of GPT-5. 5 and really what that means for our joint customers in the enterprise as they think about deploying and adopting AI. >> Thank you so much for having me. This is a really exciting day. Probably my most exciting model launch in a while. It's a really high-quality model I've had a chance to play with personally and I'm I'm excited for it to accelerate our partnership that's been ongoing for, you know, several years now. >> Oh, thank you.
We really appreciate the help in testing out the model, giving us the feedback, helping us to be better. We're really excited about the launch today. Can you tell us what Databricks is focused on right now when you think about what your customers want from enterprise AI? >> Yeah, so Databricks is a data and AI platform that works with, you know, a huge majority of the large companies across the world. And as they look to deploy agents into production, both for internal and external use cases, there's really two ingredients that matter.
There's the data and the AI. For AI, we work with partners like like you all to get our customers access to frontier intelligence. And on the data side, there's a lot of challenges that go into interacting with data when you want to deploy an AI application. Getting access to the data, making sure the data is handled securely, sifting through large volumes of structured and unstructured data. So, so where we come in is kind of where the fundamental intelligence meets the real concrete problems that enterprises are trying to apply AI to. >> Yep.
Yep. And there's a lot to think about with companies as they're thinking about bringing their data and AI together. What are you hearing from customers that they need when they're bringing AI into Databricks? >> So, we're primarily focused on two big challenges we see inside of enterprises. One is that in the same way that software engineers are significantly empowered with these newest AI coding models, companies want every employee to be able to take advantage of the benefits of intelligence.

1:39 Addressing enterprise challenges

And it turns out that's surprisingly difficult to accomplish and it's not difficult because the reasoning in the models is not good enough. In fact, the models are really really excellent reasoners. They can solve extremely difficult math and science problems. The challenge is that if you want to access the data inside of your company, that data is messy, it's spread out, you may not have access to it. A typical company might have tens of thousands of different tables or and maybe hundreds or millions of different files that contain the data that a particular worker might need to get their job done.
So, we've invested very heavily in a product called Genie. So, you should think of it like the genie in the bottle that can answer your questions. And what that's all about is getting any worker in a company that really needs to use data for their job. That could include someone in sales, it could include someone in marketing. It's really a very pervasive set of concerns. Just helping them be able to use natural language to go in and get really high quality answers to questions and things they're trying to explore in order to get their job done. >> And I think that's really important because what we see is in implementations, particularly as you think about deploying AI, it isn't just about bringing AI in, but it's rethinking the workflow.
But then when you rethink the workflow, then you have to think about where all the data comes from. And at the end of the day, 70% of the implementations are spent around the data. So, you giving your customers a way to do that in an easy way and also to bring it out beyond just technology, but to users using natural language is a huge advantage of Databricks. >> Absolutely. And we're trying to take something that might have otherwise taken weeks or months of time and just make it happen automatically for an employee so they don't need to work with someone to prepare a data set for them or make sure that the data secure or make sure that it's that it's clean and correct.
The other big area we're investing in is as companies deploy agentic applications, there's a lot of complexity in getting those from a prototyping stage to production. That falls into a few categories. It falls into sort of coding type work and coding tools. It also falls into internal productivity type workflow management and then also knowledge work, yeah. Knowledge work and then also the end products these companies are shipping.

3:35 Scaling AI to production

So, there's a set of concerns around getting access to really high-quality models like the ones that you all have, looking at security and making sure that they're observing what these agents are doing, can looking at things like cost management, you know, cost can get out of control once the use of AI tokens starts to really ramp. That's good for you guys, but also people want to manage these costs. >> That's the thing.
What enterprises need is they want to be able to use that and associate it with value, but they also need transparency and visibility so they don't get surprised. >> Exactly. And through our partnership through our AI gateway product, customers of Databricks can not only get access to these great models, but they can have those other properties like observability, quality control, cost management as they deploy maybe hundreds or thousands of applications. >> Yeah, I mean one of the things that I often talk about with companies is AI is not just a platform decision.
It's also a new operating layer and a new way of managing technology and managing your employees and the agents that you build. So, it's a whole new paradigm and you're helping companies do that. >> Absolutely. So, so our as I said, our key focus is our Genie product and this AI gateway for deploying applications, both of which are powered by OpenAI and we're super excited to be working with you guys. >> about that. So, when enterprises deploy AI into their organization, the first thing on their mind is security and control.
And how does Databricks help companies deploy such a powerful model like 5. 5, but also making sure that the guardrails and controls are there? So, safety is first. >> Yeah, this is a very important and difficult topic and it's actually one area in which I think our partnership is so synergistic for our customers. Databricks has a history for the last 15 years of building secure and trusted places for storing and managing your enterprise data.

5:07 Security and control frameworks

At the same time, OpenAI has these really amazing models that are coming out and what companies want is the ability to leverage those models safely. >> And our largely this is done in Databricks through our AI Gateway. It's basically a secure access layer between your sensitive data and the models and it can ensure all kinds of things like the agents you're deploying have as few privileges as possible with the data. They don't have a a very broad amount of scope in terms of what they can do.
You're monitoring their actions. You're able to um deal with cases where an agent appears to be accessing more things than they need to. You know, these are very classic ways of dealing with enterprise security. It's a combination of prevention and detection and when our customers use our AI Gateway they're able to get all of that. >> And they have all the observability which I think is so critical. >> Exactly. >> Well, a GPT 5.
5 is now powering Codex and I'm curious, what are you hearing from customers and feedback about the model? >> So for me the biggest improvements I personally see day-to-day are just really high-quality planning and ability to kind of interpret a complex piece of a complex knowledge task. >> Yes. >> have to do with coding or even just powering some agent application and really just fully and correctly interpret the intent of what I was trying to do and then carry that out over even maybe a very large code base with high-quality.

6:25 Performance of GPT-5.5

And that's what, you know, the holy grail is for all of these high knowledge enterprise focused models and 5. 5 so far for me personally is the best I've used. >> That's uh great to hear. It's smarter, it's capable of more complex instructions and it's really accurate and people are really loving it. We've got great feedback. Are you hearing anything from your users that you can share? >> I actually have some receipts.
I'm happy to share with you. I brought the receipts. Um and the background here is is we were able to make a preview of 5. 5 available through our AI Gateway for largely for use with coding tools. That's the people that wanted to evaluate it first and I have a few quotes. >> Oh my god. >> The names are anonymous to protect the innocent. >> innocent. >> And here's I got two quotes. Um first quote. Uh quote, "A big step up from GPT 5.

4. It found issues in a 10,000 line change pull request spanning half a dozen packages." Furthermore, I did not notice any regressions compared to 5. 4, which has been my main agentic coding model for the past month. >> Not too bad. >> Yeah, so this is this is actually pretty typical. I've heard this from many users. So, uh so next one, quote, "GPT 5. 5's planning capabilities are stronger than before. 5. 5 was able to one-shot a plan for a new coding change I was working on, including suggesting a few critical factors that other models had missed." >> So good. >> Yeah, so the you know, our our philosophy at Databricks has been for both ourselves and our customers, let the developers vote with their feet. >> Right. >> You know, let give them access to all the models and let them choose which models they think are giving them the highest yield in terms of the return on their efforts. >> Yeah, and the great feedback is that they're choosing >> Yeah, and it's great to see that many of these folks after trying 5.

5, some of whom had already kind of switched over to 5. 4, are just continuing to see major improvements. >> So great. , how did GPT 5. 5 perform across the evals for your Office QA benchmark? >> Yeah, so just to give some background, Office QA is a benchmark we built at Databricks that attempts to capture knowledge work inside of enterprises. And as I said before, knowledge work in enterprises is actually a little bit different than let's say frontier scientific reasoning or math problems or things that AI models have traditionally been steered towards. >> And I just so I'm clear on it.

8:44 Office QA benchmark results

So it's a it's an office benchmark and it's showing before and after impact of the work? >> Yeah, so so the way these benchmarks work is they might have a few hundred questions with sort of certified answers. And we have such a benchmark around office office work basically. That's a fundamentally fairly different work than is typically captured in industry benchmarks. That's why we built the benchmark because as felt there was something missing.
And then as newer models come out, they kind of run through the test and you see how well they do. >> Yeah. >> And uh at any given time there's a leaderboard with, you know, who's doing the best. So, so what makes Office QA a bit different, as I said, is it involves a lot of these fairly complex tasks around extracting data from documents and combining data across different sources and things that >> when you're working between applications, answering a question, running a process. >> Exactly.
So, it was designed exactly to capture those workloads. Because those are the ones that we at Databricks feel were uniquely able to kind of explain as an enterprise company with, you know, a huge number of customers. And um so, whenever new models come out, we sort of run them through the gauntlet and see how they do. And uh and I'm happy to announce that as of yesterday, when we did our sort of official run-through of Office QA with GPT-5.
5, that 5. 5 is now the leader state-of-the-art model for this particular workload. >> Wow. That is really exciting and that's what we had hoped for. I mean, we really see this evolution moving from coding, but also out to all knowledge work. And I think we're just getting started there. >> Yep. And there's still headroom, you know, there's headroom for OpenAI to get even better at this benchmark, and I'm sure there will be competition with others, as expected.
Um but uh but you know, we'll be keeping keeping updated of the leaderboard. And uh and this is an exciting time for OpenAI. >> OpenAI. Very exciting to hear that. Speaking of use cases, what are you hearing from customers in terms of the most impactful use cases beyond coding, but into the area of knowledge work? >> Yeah. So, first of all, um I will reiterate coding. It The impact of coding of these models on coders is just amazing.

10:53 Future of enterprise AI

And I know you you said beyond coding, but I, you know, even I'm still holding on to my roots as a software engineer, and it's just unbelievable uh the amount of leverage and the amount of efficiency gains you get. It's going to change the entire industry. >> And I also think whether you're a traditional developer or you're someone who wants to build, but maybe didn't have the skill set to do so, the idea that we can all be builders, myself included, who is not a software engineer by training.
The idea that we can all lean into a new way of thinking is really incredible. So, just in the world of the software development life cycle alone, there's still so much room to run in that regard. >> Absolutely. And I know that wasn't your question, so I'll answer your question. Um outside of coding, I mean, the two big use cases we see are are one is looking at other internal processes in a company and just just trying to automate those with AI.
And in particular are the things like the planning ability, the things we were just talking about of the latest class of models. Uh they can really uh do quite well when um when sort of confronted with existing business processes. Now, it's difficult because there's a lot of challenges around what secure How do I do with security? What data should these models have access to for these internal use cases? In some cases, business processes even have to change because the way you did something before with humans, you wouldn't do it in the same >> It's actually sometimes easier. >> Yeah, yeah, it's it's often easier.
But people need to adjust to that change. So, so whereas a coder lives in a small world where they their job was already text in text out, that's what a coder would do. They would get some specifications, they would write some code. For an an office worker, you know, their job is and the way that they interact with other people and with processes and with tools is actually much more rich. So, so the big area we see now at the frontier is is taking that type of work and then figuring out how to automate and accelerate >> Yeah, and how to simplify it.
One thing I love about working with you and the team at Databricks is that you have your industry blueprints and you also have functional blueprints and I think that's a great opportunity when you think about how we can jointly help customers in the enterprise to deploy AI. Of course, coding being the number one use case right now, but I also think what companies are looking for is guidance. Where should I go next? What is the biggest opportunity?
And I think just what your team has done um in in in creating a reference architecture around that by industry and by function is really incredible and it's a great show of how you really not just deliver the technology but help companies adopt. >> Yeah, thank you. And you know, what our strategy is we we talk with so many companies that we kind of start to see the pattern. >> Yes. >> And we start to learn what works and what doesn't work.
And we feel that the best thing we can do is share share that broadly. So, as you said, it breaks down both into sort of verticals like the way that an oil and gas company might approach the set of technologies is different than a pharmaceutical company. But it also relates to different internal functions inside of a company. So, we do see patterns. You know, marketing teams find very similar ways to leverage AI. >> Yeah. >> Um >> Data analytics in general has to be just a huge area. >> Exactly.
So, so yeah, you know, we're we're hoping to work with you all and really help accelerate the adoption really beyond coding in the next year or two. >> Yeah, I think that's like that's the spirit of our partnership and that's what we hear from our customers. It's the technology and we're so glad for your feedback on the model. We're excited about that. We're going to continue to improve. But most importantly making sure that we work in partnership to help our customers adopt the technology and we're excited about doing that with you and we're really excited about the model and we appreciate the help and the feedback on that. >> Yeah, totally.
And we're looking forward to making 5. 5 available very broadly in our AI gateway product and I'm hoping to have you know, hundreds of more quotes like that to share with you. So far we only were able to give it to a few users but the initial results are very positive. >> We're we're really pleased with the initial results. Excited about the launch and excited about the partnership. Thank you. >> Thanks for having me.

Learn more about the Databricks Data and AI platform.