The Model Mix Is Growing, and Moving In-House

Developers are increasingly using multiple models from various labs, with a growing number running on self-hosted models, according to data from OpenRouter, the company's usage dashboard, and a presentation by Revolut's AI operations leader. The trend of using a single model is waning, and the default model is now one that teams can run on their own hardware.

Customers spending over $100 a month now use an average of 6 model families, a two-fold increase from the previous year. The top 10% of customers have expanded their usage from 7 model families to 14. The usage distribution varies with spending levels: $100-$1K a month users use 5 labs, $1K-$10K users use 8 labs, and those spending over $10K use 12 labs.

Over 90% of the $10K+ group uses five or more labs. Loyalty to a single lab is rare, with fewer than 4% of major labs' users on OpenRouter using just one lab, even Anthropic, which is considered the most loyal, only has 39% of its customers' spend. The diversity in model usage is apparent within Kilo, with each agent role having its preferred model.

No single model dominates more than 30% of any role. Nvidia's Nemotron models are favored for code, plan, and debug, while review tasks see a mix of Claude Opus and GPT models. Developers use cheaper models for writing and debugging code, then switch to stronger models for error-checking. Some companies, like Revolut, are now serving self-hosted open models as the default coding model for a significant portion of their workforce.

Revolut's switch to self-hosted GLM 5.3 as the default coding model for about 2,000 engineers resulted in a 70% improvement in lead time and a 60% improvement in coding time, leading to better overall efficiency. The trend suggests a future where cheaper models handle routine tasks, while frontier models are reserved for critical steps where they provide the most value, and local AI proves effective at enterprise scale.