The performance gap between open-source and closed AI models shrank from six months in 2024 to roughly three months by 2025. If Mistral CEO Arthur Mensch is right, that gap will keep closing until model quality alone stops being a competitive advantage.
In a recent interview on the Big Technology Podcast with Alex Kantrowitz, Mensch laid out a blunt case: foundational AI models are commodities. About ten labs worldwide have the knowledge to build them, the techniques circulate freely between teams, and no one can maintain a lasting edge through architecture alone. Companies pouring billions into training runs are building assets that depreciate fast.
The Customization Bet
Mensch's alternative pitch is that the money isn't in building the biggest model. It's in helping enterprises customize smaller, specialized models on their own data. "We create models that are very easy to customize... the two things are extremely linked together," he said, describing Mistral's approach of pairing open-source models with enterprise deployment services.
This is a direct shot at the OpenAI and Anthropic playbook, where the value proposition is access to the most capable general-purpose model behind an API. Mensch is betting that enterprises - particularly in manufacturing, defense, and regulated industries - will pay for models they can own and modify rather than rent from a US provider.
Why Vertical Beats General-Purpose
The most interesting part of Mensch's argument is his rejection of the AGI framing entirely. He argues that specialized models for specific domains (materials science, aircraft design, chemistry) will create more real-world value than universal systems trying to do everything. Transfer learning between domains is limited enough that a great chemistry model doesn't help you design airplane wings.
His prediction: the next two years bring a "verticalization explosion" where smaller, cheaper, domain-specific models replace the current race toward ever-larger general systems.
There's a self-serving element here. Mistral, valued at around $6 billion, cannot match the compute budgets of Microsoft-backed OpenAI or Google DeepMind. Arguing that raw model scale doesn't matter is exactly what you'd expect from the European challenger spending less. But the technical argument about performance convergence at high compute levels (specifically around 10^26 FLOPS of training compute, the rough threshold where returns start diminishing) has support from multiple research groups.
For anyone choosing AI tools today, the practical takeaway is worth considering: if models really are converging in capability, then pricing, customization options, and data control matter more than benchmark scores. That's already playing out in the market, where Claude, GPT-4o, Gemini, and Mistral's models trade wins on different tasks rather than any one dominating across the board.