Written from the verified transcript and checked against it. Every figure links to the moment it was said.
Aaron Levie, CEO of Box, discussed the latest frontier AI models, noting that Gemini 3.x, GPT 5.x, and Opus 46-48 are all highly capable at coding and advanced reasoning, with performance differences by domain. He said the rate of improvement is not slowing, citing 2-5 point capability gains per model version. Levie dismissed the 'SaaS apocalypse' as an overreaction, arguing that systems of record like Salesforce, ServiceNow, and Workday become more valuable as agents need guardrails. He described the future of work as every individual contributor managing a fleet of agents, but emphasized that people managers and sales teams remain essential, pointing out that AI labs are hiring thousands of people. He predicted most companies will deliver 2-5 times more software within two years.
Key takeaways
Levie said frontier models are so close that for 80-90% of use cases, performance is the same across providers.
He predicted most companies will deliver 2-5 times more software within the next year or two.
Levie dismissed the 'SaaS apocalypse' as an overreaction, saying systems of record become more valuable with AI agents.
Levie said the future individual contributor will manage 3-5-10 agents constantly, but people managers will still exist.
Numbers and commitments
Figure
What it refers to
Type
At
80-90%
of use cases where frontier models perform equally
Yeah, we have our conference tomorrow. And then I got to fly right back to California. It's crazy times. Things are moving so fast that it's hard to leave for too long.
Yeah, everybody's still figuring it out. So first part, AI tech's honest review of frontier AI. You had early access of frontier models from OpenAI, Anthropic, and Google always.
And you've been posting a lot about your thoughts on all things AI on X. You have what, two million followers on X. From your experience, what feels fundamentally different in the latest generation of AI models? What are they ultimately capable of as of now?
I think that the category of models we've recently seen, the Gemini 3 plus series (3.1, 3.5), the GPT series of 5 plus (5.4, 5.5), and then Opus 46, 47, 48. That series, for all of these types of models, they've all reached a point of getting incredible at coding. They're very good at coding tasks. They're much better at advanced reasoning use cases. They're quite good for a number of specialty fields like legal, healthcare, life sciences, finance. And so what's happening is these models are obviously very good at coding, good at using tools, good at using computers, but they also have industry expertise focus areas, and they can use those tools and that expertise to execute on much longer-running tasks and work in the enterprise. So the big difference is it's no longer just asking a question like what's the weather or the history of something where it just has to look at Wikipedia and pull out information. This is much more of a task of I want you to go analyze 100 documents and pull out all the key details. It's that kind of style of work that these models are now capable of.
I think they're honestly all so close that for 80 to 90% of use cases you're going to get the same level of performance. Where they break down is more in-depth tasks. Some people will say Opus is better at front-end web development or something like GPT 5.5 is better at back-end web development. Or some model is a bit better tuned for legal work than another. So by industry and domain, each model can be a little different, mostly because of the data those models are trained on and their post-training process and what they optimize for. You're always making a trade-off when you train an AI model: what do you want it to be really good at? You only have so much time and compute. So these models have to have some degree of specialization. What we're seeing is the latest round of models, the specialization tends to be in coding-related activities. But what's incredibly exciting is that the rate of improvement is really not slowing down. Every single new version of a model, we tend to see 2, 3, 5 point improvements in capability. So the rate of progress and innovation has not slowed down at all.
Well, you have an incredible amount of talent aimed at these problems. You have a tremendous amount of capital. These companies have massive amounts of compute. They're able to buy large amounts of data sets from data labeling vendors. And I think these teams are getting better at new methodologies and techniques to make their models better. So you have a lot of great factors to improve what these models are capable of. Most of the rumors and conversation in the industry is that this is still going to be a year where we're seeing continuous breakthroughs. If you look at the model Mythos from Anthropic, it represents another step change in security capability and encoding capabilities. Once that model is broadly available, I think you'll see enterprises again see that there's all new capability available inside these AI models. There's been a lot of discussion around Mythos, especially in Japan, where governments and financial institutions are talking about it.
